9 val top_down_run : Ata.t -> Tree.t -> Tree.node -> result_set
10 val bottom_up_run : Ata.t -> Tree.t -> Compile.text_query * string -> result_set
11 val grammar_run : Ata.t -> Grammar2.t -> unit -> result_set
15 module Make (U : ResJIT.S) : S with type result_set = U.NS.t =
18 type result_set = U.NS.t;;
20 let eval_form auto s1 s2 f =
22 match Formula.expr f with
23 | Formula.False | Formula.True | Formula.Pred _ -> f, []
24 | Formula.Atom(`Left, b, q) ->
25 Formula.of_bool (b == (StateSet.mem q s1)),
26 if b && StateSet.mem q auto.topdown_marking_states then [ResJIT.LEFT q] else []
27 | Formula.Atom (`Right, b, q) ->
28 Formula.of_bool(b == (StateSet.mem q s2)),
29 if b && StateSet.mem q auto.topdown_marking_states then [ResJIT.RIGHT q] else []
30 | Formula.Atom (`Epsilon, _, _) -> assert false
32 | Formula.Or(f1, f2) ->
33 let b1, i1 = loop f1 in
34 let b2, i2 = loop f2 in
35 Formula.or_pred b1 b2, i1 @ i2
36 | Formula.And(f1, f2) ->
37 let b1, i1 = loop f1 in
38 let b2, i2 = loop f2 in
39 Formula.and_pred b1 b2, i1 @ i2
44 let eval_trans auto s1 s2 trans =
46 (fun t ((a_st, a_op, a_todo) as acc)->
47 let q, _, m, f = Transition.node t in
48 let form, ops = eval_form auto s1 s2 f in
49 match Formula.expr form with
52 (q, (if m then (ResJIT.SELF() :: ops) else ops)):: a_op,
54 | Formula.False -> acc
55 | Formula.Pred p -> a_st, a_op,
56 (p.Tree.Predicate.node, q, [(q,(if m then (ResJIT.SELF() :: ops) else ops))]) :: a_todo
58 ) trans (StateSet.empty, [], [])
65 type opcode = (t -> t -> t -> Tree.t -> Tree.node -> StateSet.t * t)
67 type t = opcode Cache.Lvl3.t
69 let dummy _ _ _ _ _ = failwith "Uninitialized L3JIT"
74 Cache.Lvl3.iteri (fun _ _ _ _ b -> if not b then incr count) a;
75 eprintf "%!L3JIT: %i used entries\n%!" !count
77 let v = Cache.Lvl3.create 1024 dummy in
78 if !Options.verbose then at_exit (fun () -> show_stats v);
81 let find t tlist s1 s2 =
83 (Uid.to_int s2.StateSet.Node.id)
84 (Uid.to_int s1.StateSet.Node.id)
85 (Uid.to_int tlist.Translist.Node.id)
87 let add t tlist s1 s2 v =
89 (Uid.to_int s2.StateSet.Node.id)
90 (Uid.to_int s1.StateSet.Node.id)
91 (Uid.to_int tlist.Translist.Node.id)
94 let compile auto trl s1 s2 =
95 let orig_s1, orig_s2 =
96 Translist.fold (fun t (a1, a2) ->
97 let _, _, _, f = Transition.node t in
98 let (_, _, fs1), (_, _, fs2) = Formula.st f in
99 (StateSet.union a1 fs1, StateSet.union a2 fs2)
100 ) trl (StateSet.empty, StateSet.empty)
102 let ns1 = StateSet.inter s1 orig_s1
103 and ns2 = StateSet.inter s2 orig_s2 in
104 let res, ops, todo = eval_trans auto ns1 ns2 trl in
105 let code, not_marking = ResJIT.compile ops in
106 let todo_code, todo_notmarking =
107 List.fold_left (fun (l, b) (p, q, o) -> let c, b' = ResJIT.compile o in
108 (p, q, c)::l, b && b')
109 ([], not_marking) todo
111 let opcode = res, code, todo_notmarking, todo_code in
114 let gen_code auto tlist s1 s2 =
115 let res, code, not_marking, todo_code = compile auto tlist s1 s2 in
117 if todo_code == [] then
118 if not_marking then begin fun empty_slot sl1 sl2 _ node ->
119 let slot1_empty = sl1 == empty_slot
120 and slot2_empty = sl2 == empty_slot in
121 if slot1_empty && slot2_empty then res,sl2
126 Array.copy empty_slot
130 U.exec sl sl1 sl2 node code;
133 else (* marking *) begin fun empty_slot sl1 sl2 _ node ->
135 if sl2 == empty_slot then
136 if sl1 == empty_slot then
137 Array.copy empty_slot
141 U.exec sl sl1 sl2 node code;
144 else (* todo != [] *)
145 begin fun empty_slot sl1 sl2 tree node ->
147 if sl2 == empty_slot then
148 if sl1 == empty_slot then
149 Array.copy empty_slot
153 U.exec sl sl1 sl2 node code;
155 (fun ares (p, q, code) ->
156 if !p tree node then begin
157 if code != ResJIT.Nil then U.exec sl sl1 sl2 node code;
160 else ares) res todo_code, sl
166 let cache_apply cache auto tlist s1 s2 =
167 let f = gen_code auto tlist s1 s2 in
168 TRACE("grammar", 2, __ "Inserting: %i, %a, %a\n%!"
169 (Uid.to_int tlist.Translist.Node.id) StateSet.print s1 StateSet.print s2);
170 add cache tlist s1 s2 f; f
173 DEFINE LOOP (t, states, ctx) = (
175 TRACE("top-down-run", 3,
176 __ "Entering node %i with loop (tag %s, context %i) with states %a\n%!"
178 (Tag.to_string (Tree.tag tree _t))
180 (StateSet.print) (states));
181 if _t == Tree.nil then nil_res
183 let tag = Tree.tag tree _t in
185 _t tag (states) (ctx) (L2JIT.find cache2 tag (states))
188 DEFINE LOOP_TAG (t, states, tag, ctx) = (
189 let _t = (t) in (* to avoid duplicating expression t *)
190 TRACE("top-down-run", 3,
191 __ "Entering node %i with loop_tag (tag %s, context %i) with states %a\n%!"
193 (Tag.to_string (tag))
195 (StateSet.print) (states));
196 if _t == Tree.nil then nil_res
199 _t (tag) (states) (ctx) (L2JIT.find cache2 (tag) (states)))
201 DEFINE LOOP(t, states, ctx) = loop (t) (states) (ctx)
202 DEFINE LOOP_TAG(t, states, tag, ctx) = loop_tag (t) (states) (ctx) (tag)
204 let top_down_run auto tree root states ctx =
205 let res_len = StateSet.max_elt auto.states + 1 in
206 let empty_slot = Array.create res_len U.NS.empty in
207 let nil_res = auto.bottom_states, empty_slot in
208 let cache3 = L3JIT.create () in
210 fun s subtree -> if subtree != U.NS.empty then
211 let r = Array.copy empty_slot in
212 r.(auto.last) <- subtree;
217 let l3jit_dispatch trl s1 s2 t sl1 sl2 =
218 let f = L3JIT.find cache3 trl s1 s2 in
219 if f == L3JIT.dummy then (L3JIT.cache_apply cache3 auto trl s1 s2) empty_slot sl1 sl2 tree t
220 else f empty_slot sl1 sl2 tree t
223 let cache2 = L2JIT.create () in
225 let rec loop t states ctx =
226 if t == Tree.nil then nil_res
228 let tag = Tree.tag tree t in
230 t tag (states) (ctx) (L2JIT.find cache2 tag (states))
231 and loop_tag t states ctx tag =
232 if t == Tree.nil then nil_res
235 t (tag) (states) (ctx) (L2JIT.find cache2 (tag) (states))
236 and l2jit_dispatch t tag states ctx opcode =
238 | L2JIT.RETURN -> nil_res
240 let opcode = L2JIT.compile cache2 auto tree tag states in
241 l2jit_dispatch t tag states ctx opcode
243 | L2JIT.LEFT (tr_list, instr) ->
245 l2jit_dispatch_instr t (Tree.closing tree t) instr
247 l3jit_dispatch tr_list res1 auto.bottom_states t slot1 empty_slot
249 | L2JIT.RIGHT (tr_list, instr) ->
251 l2jit_dispatch_instr t ctx instr
253 l3jit_dispatch tr_list auto.bottom_states res2 t empty_slot slot2
255 | L2JIT.BOTH (tr_list, instr1, instr2) ->
257 l2jit_dispatch_instr t (Tree.closing tree t) instr1
260 l2jit_dispatch_instr t ctx instr2
262 l3jit_dispatch tr_list res1 res2 t slot1 slot2
264 and l2jit_dispatch_instr t ctx instr =
266 | L2JIT.FIRST_CHILD s -> LOOP ((Tree.first_child tree t), s, ctx)
267 | L2JIT.NEXT_SIBLING s -> LOOP ((Tree.next_sibling tree t), s, ctx)
269 | L2JIT.FIRST_ELEMENT s -> LOOP ((Tree.first_element tree t), s, ctx)
270 | L2JIT.NEXT_ELEMENT s -> LOOP ((Tree.next_element tree t), s, ctx)
272 | L2JIT.TAGGED_DESCENDANT (s, tag) ->
273 LOOP_TAG ((Tree.tagged_descendant tree t tag), s, tag, ctx)
275 | L2JIT.TAGGED_FOLLOWING (s, tag) ->
276 LOOP_TAG((Tree.tagged_following_before tree t tag ctx), s, tag, ctx)
278 | L2JIT.SELECT_DESCENDANT (s, _, us) ->
279 LOOP((Tree.select_descendant tree t us), s, ctx)
281 | L2JIT.SELECT_FOLLOWING (s, pt, us) ->
282 LOOP ((Tree.select_following_before tree t us ctx), s, ctx)
284 | L2JIT.TAGGED_CHILD (s, tag) ->
285 LOOP_TAG((Tree.tagged_child tree t tag), s, tag, ctx)
287 | L2JIT.TAGGED_FOLLOWING_SIBLING (s, tag) ->
288 LOOP_TAG((Tree.tagged_following_sibling tree t tag), s, tag, ctx)
290 | L2JIT.SELECT_CHILD (s, _, us) ->
291 LOOP ((Tree.select_child tree t us), s, ctx)
293 | L2JIT.SELECT_FOLLOWING_SIBLING (s, _, us) ->
294 LOOP ((Tree.select_following_sibling tree t us), s, ctx)
296 | L2JIT.TAGGED_SUBTREE(s, tag) ->
297 mark_subtree s (U.NS.subtree_tags tree t tag)
299 | L2JIT.ELEMENT_SUBTREE(s) ->
300 mark_subtree s (U.NS.subtree_elements tree t)
302 let r = LOOP (root, states, ctx) in
303 (*L3JIT.stats err_formatter cache3; *)
306 let full_top_down_run auto states tree root =
308 top_down_run auto tree root states (Tree.closing tree root)
310 let top_down_run auto tree root =
312 let res, slot = full_top_down_run auto auto.init tree root in
314 slot.(StateSet.min_elt auto.topdown_marking_states)
317 (*** Bottom-up evaluation function **)
320 Format.fprintf fmt "{ ";
321 U.NS.iter begin fun node ->
322 Format.fprintf fmt "%a " Node.print node;
324 Format.fprintf fmt "}"
326 let slot_print fmt t =
327 Array.iteri begin fun state ns ->
328 Format.eprintf "%a -> %a\n" State.print state ns_print ns;
332 let eval_trans auto tree parent res1 res2 = assert false
334 let rec uniq = function
335 | ([] | [ _ ]) as l -> l
336 | e1 :: ((e2 :: ll) as l) -> if e1 == e2 then uniq l
337 else e1 :: e2 :: (uniq ll);;
339 let bottom_up_run auto tree (query, pat) =
340 let array = time ~msg:"Timing text query" (Tree.full_text_query query tree) pat in
341 let leaves = Array.to_list array in
342 let states = auto.states in
343 let res_len = (StateSet.max_elt states) + 1 in
344 let empty_slot = Array.create res_len U.NS.empty in
345 let nil_res = auto.bottom_states, empty_slot in
346 let cache = Cache.Lvl3.create 1024 L3JIT.dummy in
347 let rec loop_leaves l acc =
351 let res, lll = bottom_up_next node ll Tree.nil in
354 eprintf "Leftover nodes: %i\n" (List.length lll);
358 and bottom_up_next node rest stop =
359 let fs = Tree.first_child tree node in
361 if fs == Tree.nil then nil_res
362 else full_top_down_run auto states tree fs
364 move_up node res1 true rest stop
366 and move_up node res is_left rest stop =
367 if node == stop then res, rest
369 let prev_sibling = Tree.prev_sibling tree node in
370 let is_left' = prev_sibling == Tree.nil in
371 let real_parent = Tree.parent tree node in
373 if is_left' then real_parent else max (Tree.first_child tree real_parent) stop
375 (* let parent = if is_left' then Tree.parent tree node else prev_sibling in *)
376 let (s1, sl1), (s2, sl2), rest' =
377 if is_left then match rest with
378 [] -> res, nil_res, rest
380 if Tree.is_right_descendant tree node next
382 let res2, rest' = bottom_up_next next rest' node in
384 else res, nil_res, rest
388 let tag = Tree.tag tree node in
389 let id1 = Uid.to_int s1.StateSet.Node.id in
390 let id2 = Uid.to_int s2.StateSet.Node.id in
392 let code = Cache.Lvl3.find cache tag id1 id2 in
393 if code == L3JIT.dummy then
397 List.fold_left (fun acc' (labels, tr) ->
398 if labels == TagSet.any || TagSet.mem tag labels
399 then Translist.cons tr acc' else acc')
401 (Hashtbl.find auto.trans q)
406 let code = L3JIT.gen_code auto trl s1 s2 in
407 Cache.Lvl3.add cache tag id1 id2 code; code
410 let res' = code empty_slot sl1 sl2 tree node in
411 move_up parent res' is_left' rest' stop
413 let _, slot = loop_leaves leaves (nil_res) in
414 slot.(StateSet.min_elt auto.topdown_marking_states)
416 let get_trans g auto tag states =
417 StateSet.fold (fun q tr_acc ->
419 (fun ((lstates, rstates, tacc) as acc) (ts, trs) ->
420 if TagSet.mem (Tag.translate tag) ts then
421 if not (TagSet.mem Tag.attribute ts) && Grammar2.is_attribute g tag
424 let _, _, _, phi = Transition.node trs in
425 let (_,_,l), (_,_,r) = Formula.st phi in
426 (StateSet.union l lstates,
427 StateSet.union r rstates,
428 Translist.cons trs tacc)
430 tr_acc (Hashtbl.find auto.trans q)
431 ) states (StateSet.empty, StateSet.empty, Translist.nil)
434 let dispatch_param0 conf id2 y0 y1 =
436 | Grammar2.C0 | Grammar2.C2 -> Grammar2.Node0 id2
437 | Grammar2.C1 | Grammar2.C5 -> Grammar2.Node1(id2,y0)
438 | Grammar2.C3 | Grammar2.C6 -> y0
439 | Grammar2.C4 -> Grammar2.Node2(id2, y0, y1)
441 let dispatch_param1 conf id2 y0 y1 =
444 | Grammar2.C3 -> Grammar2.Node0 id2
446 | Grammar2.C6 -> Grammar2.Node1(id2, y1)
447 | _ -> Grammar2.dummy_param
449 module K_down = struct
450 type t = Grammar2.n_symbol * StateSet.t
451 let hash (x,y) = HASHINT2(Node.to_int x, Uid.to_int y.StateSet.Node.id)
452 let equal (x1,y1) (x2,y2) = x1 == x2 && y1 == y2
456 type t = Grammar2.n_symbol * StateSet.t * StateSet.t * StateSet.t
458 HASHINT4 (Node.to_int a,
459 Uid.to_int b.StateSet.Node.id,
460 Uid.to_int c.StateSet.Node.id,
461 Uid.to_int d.StateSet.Node.id)
462 let equal (a1, b1, c1, d1) (a2, b2, c2, d2) =
463 a1 == a2 && b1 == b2 && c1 == c2 && d1 == d2
468 include Hashtbl.Make(K_down)
469 let dummy = StateSet.singleton State.dummy
470 let notfound l = l.(0) == dummy && l.(1) == dummy
476 let a = [| dummy; dummy |] in
480 module UCache = Hashtbl.Make(K_up)
484 out0 : StateSet.t * U.t;
485 out1 : StateSet.t * U.t;
486 main : StateSet.t * U.t
489 { in0 = StateSet.empty;
490 in1 = StateSet.empty;
498 out0 = StateSet.empty,v;
499 out1 = StateSet.empty,v;
502 let grammar_run auto g () =
503 let dummy_leaf = Grammar2.dummy_param in
504 let dummy_set = StateSet.singleton State.dummy in
505 let res_len = (StateSet.max_elt auto.states) + 1 in
506 let empty_slot = Array.create res_len U.NS.empty in
507 let nil_res = mk_nil auto.bottom_states empty_slot in
508 let empty_res = mk_empty (StateSet.empty, empty_slot) in
509 let cache3 = L3JIT.create () in
510 let dummy2 = (StateSet.empty, StateSet.empty, Translist.nil) in
511 let cache2 = Cache.Lvl2.create 512 dummy2 in
512 let rule_counter = ref 0 in
513 let preorder_counter = ref 0 in
514 let dcache = DCache.create 1023 in
515 let ucache = UCache.create 1023 in
516 let term_array = [| StateSet.empty; StateSet.empty |] in
517 let get_trans tag states =
518 let c = Cache.Lvl2.find cache2 tag (Uid.to_int states.StateSet.Node.id) in
520 let c = get_trans g auto tag states in
522 Cache.Lvl2.add cache2 tag (Uid.to_int states.StateSet.Node.id) c;
527 let lambda = ref 0 in
528 let rec start_loop idx states =
529 TRACE("grammar", 2, __ "Node %i\n%!" (Node.to_int idx));
530 if states == dummy_set then nil_res else
531 if idx < Node.null then nil_res
533 let symbol = Grammar2.start_tag g idx in
534 let fc = Grammar2.start_first_child g idx in
535 let ns = Grammar2.start_next_sibling g fc in
536 if Grammar2.is_terminal g symbol then
537 let t = Grammar2.terminal symbol in
538 terminal_loop t states (Grammar2.Leaf (~-1,0,term_array, fc)) (Grammar2.Leaf (~-1,1,term_array, ns))
540 let nt = Grammar2.non_terminal symbol in
542 let lmbd = !lambda in
543 let y0 = (Grammar2.Leaf (lmbd,0, term_array, fc))
544 and y1 = (Grammar2.Leaf (lmbd,1, term_array, ns)) in
545 rule_loop nt states y0 y1
547 and rule_loop (t : Grammar2.n_symbol) states y0 y1 =
548 if t = Node.nil || states == dummy_set then nil_res else
549 let () = incr rule_counter in
550 if !rule_counter land 65535 == 0 then begin Gc.minor() end;
551 (* let k = (t, states) in*)
552 (* let pstates = DCache.find dcache k in
553 let notfound = DCache.notfound pstates in *)
554 let rhs = Grammar2.get_rule g t in
555 let id1 = Grammar2.get_id1 rhs in
556 let id2 = Grammar2.get_id2 rhs in
557 let conf = Grammar2.get_conf rhs in
558 (* if notfound then*)
559 let ny0 = dispatch_param0 conf id2 y0 y1 in
560 let ny1 = dispatch_param1 conf id2 y0 y1 in
561 let res = dispatch_loop id1 states ny0 ny1 in
562 (* pstates.(0) <- res.in0;
563 pstates.(1) <- res.in1; *)
565 UCache.add ucache (t, states, fst res.out0, fst res.out1)
567 let h = Hashtbl.create 7 in
568 for i = 0 to res_len - 1 do
569 Hashtbl.add h (0, i) (snd res.out0).(i);
570 Hashtbl.add h (1, i) (snd res.out1).(i);
573 main = ((fst res.main), (U.close h (snd res.main)));
577 let res0 = partial_loop y0 pstates.(0) in
578 let res1 = partial_loop y1 pstates.(1) in
579 let k2 = (t, states, fst res0.main, fst res1.main) in
582 UCache.find ucache k2
585 let ores0 = { res0 with main = fst res0.main, U.var 0 (snd res0.main) }
586 and ores1 = { res1 with main = fst res1.main, U.var 1 (snd res1.main) }
588 let res = dispatch_loop id1 states (Grammar2.Cache (0,ores0)) (Grammar2.Cache (1, ores1)) in
589 UCache.add ucache k2 res.main;
592 let h = Hashtbl.create 7 in
593 for i = 0 to res_len - 1 do
594 Hashtbl.add h (0, i) (snd res0.main).(i);
595 Hashtbl.add h (1, i) (snd res1.main).(i);
601 main = s, U.close h r;
604 and dispatch_loop id1 states ny0 ny1 =
605 if Grammar2.is_non_terminal g id1 then
606 rule_loop (Grammar2.non_terminal id1) states ny0 ny1
608 terminal_loop (Grammar2.terminal id1) states ny0 ny1
610 and terminal_loop (symbol : Grammar2.t_symbol) states y0 y1 =
612 if symbol == Grammar2.nil_symbol || symbol = Node.nil || states == dummy_set then nil_res else begin
613 let tag = Grammar2.tag symbol in
614 let lst, rst, trans = get_trans tag states in
615 let res0 = partial_loop y0 lst in
616 let res1 = partial_loop y1 rst in
617 let s1, slot1 = res0.main
618 and s2, slot2 = res1.main in
619 let opcode = L3JIT.find cache3 trans s1 s2 in
620 let node = Node.of_int !preorder_counter in
621 incr preorder_counter;
623 if opcode == L3JIT.dummy then
624 (L3JIT.cache_apply cache3 auto trans s1 s2) empty_slot slot1 slot2 (Obj.magic ()) node
626 opcode empty_slot slot1 slot2 (Obj.magic()) (node)
635 and partial_loop l states =
636 if l == dummy_leaf then nil_res else
638 | Grammar2.Cache (_, r) -> r
639 | Grammar2.Leaf (_,_, _, id) -> start_loop id states
640 | Grammar2.Node0 id ->
641 if (Grammar2.terminal id) == Grammar2.nil_symbol then nil_res
643 rule_loop (Grammar2.non_terminal id) states dummy_leaf dummy_leaf
645 | Grammar2.Node1 (id, y0) ->
646 rule_loop (Grammar2.non_terminal id) states y0 dummy_leaf
647 | Grammar2.Node2 (id, y0, y1) ->
648 if Grammar2.is_terminal g id then
649 terminal_loop (Grammar2.terminal id) states y0 y1
651 rule_loop (Grammar2.non_terminal id) states y0 y1
654 let (_, slot) = (start_loop (Node.null) auto.init).main in
655 slot.(StateSet.min_elt auto.topdown_marking_states)