src/HOL/Tools/Sledgehammer/sledgehammer_compress.ML
changeset 51130 76d68444cd59
parent 51128 0021ea861129
child 51147 020a6812a204
--- /dev/null	Thu Jan 01 00:00:00 1970 +0000
+++ b/src/HOL/Tools/Sledgehammer/sledgehammer_compress.ML	Thu Feb 14 22:49:22 2013 +0100
@@ -0,0 +1,252 @@
+(*  Title:      HOL/Tools/Sledgehammer/sledgehammer_compress.ML
+    Author:     Jasmin Blanchette, TU Muenchen
+    Author:     Steffen Juilf Smolka, TU Muenchen
+
+Compression of reconstructed isar proofs.
+*)
+
+signature SLEDGEHAMMER_COMPRESS =
+sig
+  type isar_step = Sledgehammer_Proof.isar_step
+  type preplay_time = Sledgehammer_Preplay.preplay_time
+  val compress_proof :
+    bool -> Proof.context -> string -> string -> bool -> Time.time option
+    -> real -> isar_step list -> isar_step list * (bool * preplay_time)
+end
+
+structure Sledgehammer_Compress : SLEDGEHAMMER_COMPRESS =
+struct
+
+open Sledgehammer_Util
+open Sledgehammer_Proof
+open Sledgehammer_Preplay
+
+(* Parameters *)
+val merge_timeout_slack = 1.2
+
+(* Data structures, orders *)
+val label_ord = prod_ord int_ord fast_string_ord o pairself swap
+structure Label_Table = Table(
+  type key = label
+  val ord = label_ord)
+
+(* clean vector interface *)
+fun get i v = Vector.sub (v, i)
+fun replace x i v = Vector.update (v, i, x)
+fun update f i v = replace (get i v |> f) i v
+fun v_map_index f v = Vector.foldr (op::) nil v |> map_index f |> Vector.fromList
+fun v_fold_index f v s =
+  Vector.foldl (fn (x, (i, s)) => (i+1, f (i, x) s)) (0, s) v |> snd
+
+(* Queue interface to table *)
+fun pop tab key =
+  let val v = hd (Inttab.lookup_list tab key) in
+    (v, Inttab.remove_list (op =) (key, v) tab)
+  end
+fun pop_max tab = pop tab (the (Inttab.max_key tab))
+fun add_list tab xs = fold (Inttab.insert_list (op =)) xs tab
+
+(* Main function for compresing proofs *)
+fun compress_proof debug ctxt type_enc lam_trans preplay preplay_timeout
+                   isar_compress proof =
+  let
+    (* 60 seconds seems like a good interpreation of "no timeout" *)
+    val preplay_timeout = preplay_timeout |> the_default (seconds 60.0)
+
+    (* handle metis preplay fail *)
+    local
+      open Unsynchronized
+      val metis_fail = ref false
+    in
+      fun handle_metis_fail try_metis () =
+        try_metis () handle exn =>
+          (if Exn.is_interrupt exn orelse debug then reraise exn
+           else metis_fail := true; some_preplay_time)
+      fun get_time lazy_time =
+        if !metis_fail andalso not (Lazy.is_finished lazy_time)
+          then some_preplay_time
+          else Lazy.force lazy_time
+      val metis_fail = fn () => !metis_fail
+    end
+
+    (* compress proof on top level - do not compress case splits *)
+    fun compress_top_level on_top_level ctxt proof =
+    let
+      (* proof vector *)
+      val proof_vect = proof |> map SOME |> Vector.fromList
+      val n = Vector.length proof_vect
+      val n_metis = metis_steps_top_level proof
+      val target_n_metis = Real.fromInt n_metis / isar_compress |> Real.round
+
+      (* table for mapping from (top-level-)label to proof position *)
+      fun update_table (i, Assume (l, _)) = Label_Table.update_new (l, i)
+        | update_table (i, Obtain (_, _, l, _, _)) = Label_Table.update_new (l, i)
+        | update_table (i, Prove (_, l, _, _)) = Label_Table.update_new (l, i)
+        | update_table _ = I
+      val label_index_table = fold_index update_table proof Label_Table.empty
+      val lookup_indices = map_filter (Label_Table.lookup label_index_table)
+
+      (* proof references *)
+      fun refs (Obtain (_, _, _, _, By_Metis (lfs, _))) = lookup_indices lfs
+        | refs (Prove (_, _, _, By_Metis (lfs, _))) = lookup_indices lfs
+        | refs (Prove (_, _, _, Case_Split (cases, (lfs, _)))) =
+          lookup_indices lfs @ maps (maps refs) cases
+        | refs (Prove (_, _, _, Subblock proof)) = maps refs proof
+        | refs _ = []
+      val refed_by_vect =
+        Vector.tabulate (n, (fn _ => []))
+        |> fold_index (fn (i, step) => fold (update (cons i)) (refs step)) proof
+        |> Vector.map rev (* after rev, indices are sorted in ascending order *)
+
+      (* candidates for elimination, use table as priority queue (greedy
+         algorithm) *)
+      fun add_if_cand proof_vect (i, [j]) =
+          (case (the (get i proof_vect), the (get j proof_vect)) of
+            (Prove (_, _, t, By_Metis _), Prove (_, _, _, By_Metis _)) =>
+              cons (Term.size_of_term t, i)
+          | (Prove (_, _, t, By_Metis _), Obtain (_, _, _, _, By_Metis _)) =>
+              cons (Term.size_of_term t, i)
+          | _ => I)
+        | add_if_cand _ _ = I
+      val cand_tab =
+        v_fold_index (add_if_cand proof_vect) refed_by_vect []
+        |> Inttab.make_list
+
+      (* cache metis preplay times in lazy time vector *)
+      val metis_time =
+        v_map_index
+          (if not preplay then K (zero_preplay_time) #> Lazy.value
+           else
+             apsnd the (* step *)
+             #> apfst (fn i => try (get (i-1) #> the) proof_vect) (* succedent *)
+             #> try_metis debug type_enc lam_trans ctxt preplay_timeout
+             #> handle_metis_fail
+             #> Lazy.lazy)
+          proof_vect
+
+      fun sum_up_time lazy_time_vector =
+        Vector.foldl
+          (apfst get_time #> uncurry add_preplay_time)
+          zero_preplay_time lazy_time_vector
+
+      (* Merging *)
+      fun merge (Prove (_, label1, _, By_Metis (lfs1, gfs1))) step2 =
+          let
+            val (step_constructor, lfs2, gfs2) =
+              (case step2 of
+                (Prove (qs2, label2, t, By_Metis (lfs2, gfs2))) =>
+                  (fn by => Prove (qs2, label2, t, by), lfs2, gfs2)
+              | (Obtain (qs2, xs, label2, t, By_Metis (lfs2, gfs2))) =>
+                  (fn by => Obtain (qs2, xs, label2, t, by), lfs2, gfs2)
+              | _ => error "sledgehammer_compress: unmergeable Isar steps" )
+            val lfs = remove (op =) label1 lfs2 |> union (op =) lfs1
+            val gfs = union (op =) gfs1 gfs2
+          in step_constructor (By_Metis (lfs, gfs)) end
+        | merge _ _ = error "sledgehammer_compress: unmergeable Isar steps"
+
+      fun try_merge metis_time (s1, i) (s2, j) =
+        if not preplay then (merge s1 s2 |> SOME, metis_time)
+        else
+          (case get i metis_time |> Lazy.force of
+            (true, _) => (NONE, metis_time)
+          | (_, t1) =>
+            (case get j metis_time |> Lazy.force of
+              (true, _) => (NONE, metis_time)
+            | (_, t2) =>
+              let
+                val s12 = merge s1 s2
+                val timeout = time_mult merge_timeout_slack (Time.+(t1, t2))
+              in
+                case try_metis_quietly debug type_enc lam_trans ctxt timeout
+                (NONE, s12) () of
+                  (true, _) => (NONE, metis_time)
+                | exact_time =>
+                  (SOME s12, metis_time
+                             |> replace (zero_preplay_time |> Lazy.value) i
+                             |> replace (Lazy.value exact_time) j)
+
+              end))
+
+      fun merge_steps metis_time proof_vect refed_by cand_tab n' n_metis' =
+        if Inttab.is_empty cand_tab
+          orelse n_metis' <= target_n_metis
+          orelse (on_top_level andalso n'<3)
+        then
+          (Vector.foldr
+             (fn (NONE, proof) => proof | (SOME s, proof) => s :: proof)
+             [] proof_vect,
+           sum_up_time metis_time)
+        else
+          let
+            val (i, cand_tab) = pop_max cand_tab
+            val j = get i refed_by |> the_single
+            val s1 = get i proof_vect |> the
+            val s2 = get j proof_vect |> the
+          in
+            case try_merge metis_time (s1, i) (s2, j) of
+              (NONE, metis_time) =>
+              merge_steps metis_time proof_vect refed_by cand_tab n' n_metis'
+            | (s, metis_time) =>
+            let
+              val refs = refs s1
+              val refed_by = refed_by |> fold
+                (update (Ord_List.remove int_ord i #> Ord_List.insert int_ord j)) refs
+              val new_candidates =
+                fold (add_if_cand proof_vect)
+                  (map (fn i => (i, get i refed_by)) refs) []
+              val cand_tab = add_list cand_tab new_candidates
+              val proof_vect = proof_vect |> replace NONE i |> replace s j
+            in
+              merge_steps metis_time proof_vect refed_by cand_tab (n' - 1)
+                          (n_metis' - 1)
+            end
+          end
+    in
+      merge_steps metis_time proof_vect refed_by_vect cand_tab n n_metis
+    end
+
+    fun do_proof on_top_level ctxt proof =
+      let
+        (* Enrich context with top-level facts *)
+        val thy = Proof_Context.theory_of ctxt
+        (* TODO: add Skolem variables to context? *)
+        fun enrich_with_fact l t =
+          Proof_Context.put_thms false
+            (string_for_label l, SOME [Skip_Proof.make_thm thy t])
+        fun enrich_with_step (Assume (l, t)) = enrich_with_fact l t
+          | enrich_with_step (Obtain (_, _, l, t, _)) = enrich_with_fact l t
+          | enrich_with_step (Prove (_, l, t, _)) = enrich_with_fact l t
+          | enrich_with_step _ = I
+        val rich_ctxt = fold enrich_with_step proof ctxt
+
+        (* compress subproofs (case_splits and subblocks) and top-levl *)
+        val ((proof, top_level_time), lower_level_time) =
+          proof |> do_subproof rich_ctxt
+                |>> compress_top_level on_top_level rich_ctxt
+      in
+        (proof, add_preplay_time lower_level_time top_level_time)
+      end
+
+    and do_subproof ctxt proof =
+      let
+        fun compress_each_and_collect_time compress candidates =
+          let fun f_m cand time = compress cand ||> add_preplay_time time
+          in fold_map f_m candidates zero_preplay_time end
+        val compress_subproof =
+          compress_each_and_collect_time (do_proof false ctxt)
+        fun compress (Prove (qs, l, t, Case_Split (cases, facts))) =
+            let val (cases, time) = compress_subproof cases
+            in (Prove (qs, l, t, Case_Split (cases, facts)), time) end
+          | compress (Prove (qs, l, t, Subblock proof)) =
+            let val ([proof], time) = compress_subproof [proof]
+            in (Prove (qs, l, t, Subblock proof), time) end
+          | compress step = (step, zero_preplay_time)
+      in
+        compress_each_and_collect_time compress proof
+      end
+  in
+    do_proof true ctxt proof
+    |> apsnd (pair (metis_fail ()))
+  end
+
+end