DFA state compression: 16-bit wide and sherman co-exist

src/grey.cpp

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2015-2017, Intel Corporation
+ * Copyright (c) 2015-2018, Intel Corporation
  *
  * Redistribution and use in source and binary forms, with or without
  * modification, are permitted provided that the following conditions are met:
@@ -82,6 +82,7 @@ Grey::Grey(void) :
                    onlyOneOutfix(false),
                    allowShermanStates(true),
                    allowMcClellan8(true),
+                   allowWideStates(true), // enable wide state for McClellan8
                    highlanderPruneDFA(true),
                    minimizeDFA(true),
                    accelerateDFA(true),
@@ -251,6 +252,7 @@ void applyGreyOverrides(Grey *g, const string &s) {
         G_UPDATE(onlyOneOutfix);
         G_UPDATE(allowShermanStates);
         G_UPDATE(allowMcClellan8);
+        G_UPDATE(allowWideStates);
         G_UPDATE(highlanderPruneDFA);
         G_UPDATE(minimizeDFA);
         G_UPDATE(accelerateDFA);

src/grey.h

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2015-2017, Intel Corporation
+ * Copyright (c) 2015-2018, Intel Corporation
  *
  * Redistribution and use in source and binary forms, with or without
  * modification, are permitted provided that the following conditions are met:
@@ -87,6 +87,7 @@ struct Grey {
 
     bool allowShermanStates;
     bool allowMcClellan8;
+    bool allowWideStates; // enable wide state for McClellan8
     bool highlanderPruneDFA;
     bool minimizeDFA;
 

src/nfa/accel_dfa_build_strat.h

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2015-2017, Intel Corporation
+ * Copyright (c) 2015-2018, Intel Corporation
  *
  * Redistribution and use in source and binary forms, with or without
  * modification, are permitted provided that the following conditions are met:
@@ -40,6 +40,11 @@ namespace ue2 {
 
 class ReportManager;
 struct Grey;
+enum DfaType {
+    McClellan,
+    Sheng,
+    Gough
+};
 
 class accel_dfa_build_strat : public dfa_build_strat {
 public:
@@ -53,6 +58,8 @@ public:
     virtual void buildAccel(dstate_id_t this_idx, const AccelScheme &info,
                             void *accel_out);
     virtual std::map<dstate_id_t, AccelScheme> getAccelInfo(const Grey &grey);
+    virtual DfaType getType() const = 0;
+
 private:
     bool only_accel_init;
 };

src/nfa/goughcompile.cpp

@@ -91,6 +91,7 @@ public:
     void buildAccel(dstate_id_t this_idx, const AccelScheme &info,
                     void *accel_out) override;
     u32 max_allowed_offset_accel() const override { return 0; }
+    DfaType getType() const override { return Gough; }
 
     raw_som_dfa &rdfa;
     const GoughGraph ≫

src/nfa/mcclellan.c

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2015-2017, Intel Corporation
+ * Copyright (c) 2015-2018, Intel Corporation
  *
  * Redistribution and use in source and binary forms, with or without
  * modification, are permitted provided that the following conditions are met:
@@ -167,9 +167,68 @@ u32 doNormal16(const struct mcclellan *m, const u8 **c_inout, const u8 *end,
 }
 
 static really_inline
-char mcclellanExec16_i(const struct mcclellan *m, u32 *state, const u8 *buf,
+u32 doNormalWide16(const struct mcclellan *m, const u8 **c_inout,
-                       size_t len, u64a offAdj, NfaCallback cb, void *ctxt,
+                   const u8 *end, u32 s, char *qstate, u16 *offset,
-                       char single, const u8 **c_final, enum MatchMode mode) {
+                   char do_accel, enum MatchMode mode) {
+    const u8 *c = *c_inout;
+
+    u32 wide_limit = m->wide_limit;
+    const char *wide_base
+        = (const char *)m - sizeof(struct NFA) + m->wide_offset;
+
+    const u16 *succ_table
+        = (const u16 *)((const char *)m + sizeof(struct mcclellan));
+    assert(ISALIGNED_N(succ_table, 2));
+    u32 sherman_base = m->sherman_limit;
+    const char *sherman_base_offset
+        = (const char *)m - sizeof(struct NFA) + m->sherman_offset;
+    u32 as = m->alphaShift;
+
+    s &= STATE_MASK;
+
+    while (c < end && s) {
+        u8 cprime = m->remap[*c];
+        DEBUG_PRINTF("c: %02hhx '%c' cp:%02hhx (s=%u) &c: %p\n", *c,
+                     ourisprint(*c) ? *c : '?', cprime, s, c);
+
+        if (unlikely(s >= wide_limit)) {
+            const char *wide_entry
+                = findWideEntry16(m, wide_base, wide_limit, s);
+            DEBUG_PRINTF("doing wide head (%u)\n", s);
+            s = doWide16(wide_entry, &c, end, m->remap, (u16 *)&s, qstate,
+                         offset);
+        } else if (s >= sherman_base) {
+            const char *sherman_state
+                = findShermanState(m, sherman_base_offset, sherman_base, s);
+            DEBUG_PRINTF("doing sherman (%u)\n", s);
+            s = doSherman16(sherman_state, cprime, succ_table, as);
+        } else {
+            DEBUG_PRINTF("doing normal\n");
+            s = succ_table[(s << as) + cprime];
+        }
+
+        DEBUG_PRINTF("s: %u (%u)\n", s, s & STATE_MASK);
+        c++;
+
+        if (do_accel && (s & ACCEL_FLAG)) {
+            break;
+        }
+        if (mode != NO_MATCHES && (s & ACCEPT_FLAG)) {
+            break;
+        }
+
+        s &= STATE_MASK;
+    }
+
+    *c_inout = c;
+    return s;
+}
+
+static really_inline
+char mcclellanExec16_i(const struct mcclellan *m, u32 *state, char *qstate,
+                       const u8 *buf, size_t len, u64a offAdj, NfaCallback cb,
+                       void *ctxt, char single, const u8 **c_final,
+                       enum MatchMode mode) {
     assert(ISALIGNED_N(state, 2));
     if (!len) {
         if (mode == STOP_AT_MATCH) {
@@ -179,6 +238,7 @@ char mcclellanExec16_i(const struct mcclellan *m, u32 *state, const u8 *buf,
     }
 
     u32 s = *state;
+    u16 offset = 0;
     const u8 *c = buf;
     const u8 *c_end = buf + len;
     const struct mstate_aux *aux
@@ -207,7 +267,12 @@ without_accel:
             goto exit;
         }
 
-        s = doNormal16(m, &c, min_accel_offset, s, 0, mode);
+        if (unlikely(m->has_wide)) {
+            s = doNormalWide16(m, &c, min_accel_offset, s, qstate, &offset, 0,
+                               mode);
+        } else {
+            s = doNormal16(m, &c, min_accel_offset, s, 0, mode);
+        }
 
         if (mode != NO_MATCHES && (s & ACCEPT_FLAG)) {
             if (mode == STOP_AT_MATCH) {
@@ -259,7 +324,11 @@ with_accel:
             }
         }
 
-        s = doNormal16(m, &c, c_end, s, 1, mode);
+        if (unlikely(m->has_wide)) {
+            s = doNormalWide16(m, &c, c_end, s, qstate, &offset, 1, mode);
+        } else {
+            s = doNormal16(m, &c, c_end, s, 1, mode);
+        }
 
         if (mode != NO_MATCHES && (s & ACCEPT_FLAG)) {
             if (mode == STOP_AT_MATCH) {
@@ -297,44 +366,47 @@ exit:
 }
 
 static never_inline
-char mcclellanExec16_i_cb(const struct mcclellan *m, u32 *state, const u8 *buf,
+char mcclellanExec16_i_cb(const struct mcclellan *m, u32 *state, char *qstate,
-                          size_t len, u64a offAdj, NfaCallback cb, void *ctxt,
+                          const u8 *buf, size_t len, u64a offAdj,
-                          char single, const u8 **final_point) {
+                          NfaCallback cb, void *ctxt, char single,
-    return mcclellanExec16_i(m, state, buf, len, offAdj, cb, ctxt, single,
+                          const u8 **final_point) {
-                             final_point, CALLBACK_OUTPUT);
+    return mcclellanExec16_i(m, state, qstate, buf, len, offAdj, cb, ctxt,
+                             single, final_point, CALLBACK_OUTPUT);
 }
 
 static never_inline
-char mcclellanExec16_i_sam(const struct mcclellan *m, u32 *state, const u8 *buf,
+char mcclellanExec16_i_sam(const struct mcclellan *m, u32 *state, char *qstate,
-                           size_t len, u64a offAdj, NfaCallback cb, void *ctxt,
+                           const u8 *buf, size_t len, u64a offAdj,
-                           char single, const u8 **final_point) {
+                           NfaCallback cb, void *ctxt, char single,
-    return mcclellanExec16_i(m, state, buf, len, offAdj, cb, ctxt, single,
+                           const u8 **final_point) {
-                             final_point, STOP_AT_MATCH);
+    return mcclellanExec16_i(m, state, qstate, buf, len, offAdj, cb, ctxt,
+                             single, final_point, STOP_AT_MATCH);
 }
 
 static never_inline
-char mcclellanExec16_i_nm(const struct mcclellan *m, u32 *state, const u8 *buf,
+char mcclellanExec16_i_nm(const struct mcclellan *m, u32 *state, char *qstate,
-                          size_t len, u64a offAdj, NfaCallback cb, void *ctxt,
+                          const u8 *buf, size_t len, u64a offAdj,
-                          char single, const u8 **final_point) {
+                          NfaCallback cb, void *ctxt, char single,
-    return mcclellanExec16_i(m, state, buf, len, offAdj, cb, ctxt, single,
+                          const u8 **final_point) {
-                             final_point, NO_MATCHES);
+    return mcclellanExec16_i(m, state, qstate, buf, len, offAdj, cb, ctxt,
+                             single, final_point, NO_MATCHES);
 }
 
 static really_inline
-char mcclellanExec16_i_ni(const struct mcclellan *m, u32 *state, const u8 *buf,
+char mcclellanExec16_i_ni(const struct mcclellan *m, u32 *state, char *qstate,
-                          size_t len, u64a offAdj, NfaCallback cb, void *ctxt,
+                          const u8 *buf, size_t len, u64a offAdj,
-                          char single, const u8 **final_point,
+                          NfaCallback cb, void *ctxt, char single,
-                          enum MatchMode mode) {
+                          const u8 **final_point, enum MatchMode mode) {
     if (mode == CALLBACK_OUTPUT) {
-        return mcclellanExec16_i_cb(m, state, buf, len, offAdj, cb, ctxt,
+        return mcclellanExec16_i_cb(m, state, qstate, buf, len, offAdj, cb,
-                                    single, final_point);
+                                    ctxt, single, final_point);
     } else if (mode == STOP_AT_MATCH) {
-        return mcclellanExec16_i_sam(m, state, buf, len, offAdj, cb, ctxt,
+        return mcclellanExec16_i_sam(m, state, qstate, buf, len, offAdj, cb,
-                                     single, final_point);
+                                     ctxt, single, final_point);
     } else {
         assert(mode == NO_MATCHES);
-        return mcclellanExec16_i_nm(m, state, buf, len, offAdj, cb, ctxt,
+        return mcclellanExec16_i_nm(m, state, qstate, buf, len, offAdj, cb,
-                                    single, final_point);
+                                    ctxt, single, final_point);
     }
 }
 
@@ -540,6 +612,10 @@ char mcclellanCheckEOD(const struct NFA *nfa, u32 s, u64a offset,
     const struct mcclellan *m = getImplNfa(nfa);
     const struct mstate_aux *aux = get_aux(m, s);
 
+    if (m->has_wide == 1 && s >= m->wide_limit) {
+        return MO_CONTINUE_MATCHING;
+    }
+
     if (!aux->accept_eod) {
         return MO_CONTINUE_MATCHING;
     }
@@ -612,9 +688,9 @@ char nfaExecMcClellan16_Q2i(const struct NFA *n, u64a offset, const u8 *buffer,
 
         /* do main buffer region */
         const u8 *final_look;
-        char rv = mcclellanExec16_i_ni(m, &s, cur_buf + sp, local_ep - sp,
+        char rv = mcclellanExec16_i_ni(m, &s, q->state, cur_buf + sp,
-                                      offset + sp, cb, context, single,
+                                       local_ep - sp, offset + sp, cb, context,
-                                      &final_look, mode);
+                                       single, &final_look, mode);
         if (rv == MO_DEAD) {
             *(u16 *)q->state = 0;
             return MO_DEAD;
@@ -684,12 +760,16 @@ char nfaExecMcClellan16_Bi(const struct NFA *n, u64a offset, const u8 *buffer,
     const struct mcclellan *m = getImplNfa(n);
     u32 s = m->start_anchored;
 
-    if (mcclellanExec16_i(m, &s, buffer, length, offset, cb, context, single,
+    if (mcclellanExec16_i(m, &s, NULL, buffer, length, offset, cb, context,
-                          NULL, CALLBACK_OUTPUT)
+                          single, NULL, CALLBACK_OUTPUT)
         == MO_DEAD) {
         return s ? MO_ALIVE : MO_DEAD;
     }
 
+    if (m->has_wide == 1 && s >= m->wide_limit) {
+        return MO_ALIVE;
+    }
+
     const struct mstate_aux *aux = get_aux(m, s);
 
     if (aux->accept_eod) {
@@ -768,6 +848,7 @@ char nfaExecMcClellan8_Q2i(const struct NFA *n, u64a offset, const u8 *buffer,
         char rv = mcclellanExec8_i_ni(m, &s, cur_buf + sp, local_ep - sp,
                                      offset + sp, cb, context, single,
                                      &final_look, mode);
+
         if (rv == MO_HALT_MATCHING) {
             *(u8 *)q->state = 0;
             return MO_DEAD;
@@ -1016,7 +1097,8 @@ char nfaExecMcClellan16_inAccept(const struct NFA *n, ReportID report,
     u16 s = *(u16 *)q->state;
     DEBUG_PRINTF("checking accepts for %hu\n", s);
 
-    return mcclellanHasAccept(m, get_aux(m, s), report);
+    return (m->has_wide == 1 && s >= m->wide_limit) ?
+                0 : mcclellanHasAccept(m, get_aux(m, s), report);
 }
 
 char nfaExecMcClellan16_inAnyAccept(const struct NFA *n, struct mq *q) {
@@ -1026,7 +1108,8 @@ char nfaExecMcClellan16_inAnyAccept(const struct NFA *n, struct mq *q) {
     u16 s = *(u16 *)q->state;
     DEBUG_PRINTF("checking accepts for %hu\n", s);
 
-    return !!get_aux(m, s)->accept;
+    return (m->has_wide == 1 && s >= m->wide_limit) ?
+                0 : !!get_aux(m, s)->accept;
 }
 
 char nfaExecMcClellan8_Q2(const struct NFA *n, struct mq *q, s64a end) {
@@ -1111,6 +1194,12 @@ char nfaExecMcClellan16_initCompressedState(const struct NFA *nfa, u64a offset,
                                             void *state, UNUSED u8 key) {
     const struct mcclellan *m = getImplNfa(nfa);
     u16 s = offset ? m->start_floating : m->start_anchored;
+
+    // new byte
+    if (m->has_wide) {
+        *((u16 *)state + 1) = 0;
+    }
+
     if (s) {
         unaligned_store_u16(state, s);
         return 1;
@@ -1140,14 +1229,24 @@ void nfaExecMcClellan16_SimpStream(const struct NFA *nfa, char *state,
                                    const u8 *buf, char top, size_t start_off,
                                    size_t len, NfaCallback cb, void *ctxt) {
     const struct mcclellan *m = getImplNfa(nfa);
+    u32 s;
 
-    u32 s = top ? m->start_anchored : unaligned_load_u16(state);
+    if (top) {
+        s = m->start_anchored;
+
+        // new byte
+        if (m->has_wide) {
+            *((u16 *)state + 1) = 0;
+        }
+    } else {
+        s = unaligned_load_u16(state);
+    }
 
     if (m->flags & MCCLELLAN_FLAG_SINGLE) {
-        mcclellanExec16_i(m, &s, buf + start_off, len - start_off,
+        mcclellanExec16_i(m, &s, state, buf + start_off, len - start_off,
                           start_off, cb, ctxt, 1, NULL, CALLBACK_OUTPUT);
     } else {
-        mcclellanExec16_i(m, &s, buf + start_off, len - start_off,
+        mcclellanExec16_i(m, &s, state, buf + start_off, len - start_off,
                           start_off, cb, ctxt, 0, NULL, CALLBACK_OUTPUT);
     }
 
@@ -1178,9 +1277,16 @@ char nfaExecMcClellan8_queueInitState(UNUSED const struct NFA *nfa,
 
 char nfaExecMcClellan16_queueInitState(UNUSED const struct NFA *nfa,
                                        struct mq *q) {
-    assert(nfa->scratchStateSize == 2);
+    const struct mcclellan *m = getImplNfa(nfa);
+    assert(m->has_wide == 1 ? nfa->scratchStateSize == 4
+                            : nfa->scratchStateSize == 2);
     assert(ISALIGNED_N(q->state, 2));
     *(u16 *)q->state = 0;
+
+    // new byte
+    if (m->has_wide) {
+        *((u16 *)q->state + 1) = 0;
+    }
     return 0;
 }
 
@@ -1206,21 +1312,39 @@ char nfaExecMcClellan8_expandState(UNUSED const struct NFA *nfa, void *dest,
 char nfaExecMcClellan16_queueCompressState(UNUSED const struct NFA *nfa,
                                            const struct mq *q,
                                            UNUSED s64a loc) {
+    const struct mcclellan *m = getImplNfa(nfa);
     void *dest = q->streamState;
     const void *src = q->state;
-    assert(nfa->scratchStateSize == 2);
+    assert(m->has_wide == 1 ? nfa->scratchStateSize == 4
-    assert(nfa->streamStateSize == 2);
+                            : nfa->scratchStateSize == 2);
+    assert(m->has_wide == 1 ? nfa->streamStateSize == 4
+                            : nfa->streamStateSize == 2);
+
     assert(ISALIGNED_N(src, 2));
     unaligned_store_u16(dest, *(const u16 *)(src));
+
+    // new byte
+    if (m->has_wide) {
+        *((u16 *)dest + 1) = *((const u16 *)src + 1);
+    }
     return 0;
 }
 
 char nfaExecMcClellan16_expandState(UNUSED const struct NFA *nfa, void *dest,
                                     const void *src, UNUSED u64a offset,
                                     UNUSED u8 key) {
-    assert(nfa->scratchStateSize == 2);
+    const struct mcclellan *m = getImplNfa(nfa);
-    assert(nfa->streamStateSize == 2);
+    assert(m->has_wide == 1 ? nfa->scratchStateSize == 4
+                            : nfa->scratchStateSize == 2);
+    assert(m->has_wide == 1 ? nfa->streamStateSize == 4
+                            : nfa->streamStateSize == 2);
+
     assert(ISALIGNED_N(dest, 2));
     *(u16 *)dest = unaligned_load_u16(src);
+
+    // new byte
+    if (m->has_wide) {
+        *((u16 *)dest + 1) = *((const u16 *)src + 1);
+    }
     return 0;
 }

src/nfa/mcclellan_common_impl.h

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2015-2016, Intel Corporation
+ * Copyright (c) 2015-2018, Intel Corporation
  *
  * Redistribution and use in source and binary forms, with or without
  * modification, are permitted provided that the following conditions are met:
@@ -82,3 +82,108 @@ u32 doSherman16(const char *sherman_state, u8 cprime, const u16 *succ_table,
     u32 daddy = *(const u16 *)(sherman_state + SHERMAN_DADDY_OFFSET);
     return succ_table[(daddy << as) + cprime];
 }
+
+static really_inline
+u16 doWide16(const char *wide_entry, const u8 **c_inout, const u8 *end,
+             const u8 *remap, const u16 *s, char *qstate, u16 *offset) {
+    // Internal relative offset after the last visit of the wide state.
+    if (qstate != NULL) { // stream mode
+        *offset = *(const u16 *)(qstate + 2);
+    }
+
+    u8 successful = 0;
+    const u8 *c = *c_inout;
+    u32 len_c = end - c;
+
+    u16 width = *(const u16 *)(wide_entry + WIDE_WIDTH_OFFSET);
+    assert(width >= 8);
+    const u8 *symbols = (const u8 *)(wide_entry + WIDE_SYMBOL_OFFSET16);
+    const u16 *trans = (const u16 *)(wide_entry +
+                                     WIDE_TRANSITION_OFFSET16(width));
+
+    assert(*offset < width);
+    u16 len_w = width - *offset;
+    const u8 *sym = symbols + *offset;
+
+    char tmp[16];
+    u16 pos = 0;
+
+    if (*offset == 0 && remap[*c] != *sym) {
+        goto normal;
+    }
+
+    // both in (16, +oo).
+    while (len_w >= 16 && len_c >= 16) {
+        m128 str_w = loadu128(sym);
+        for (size_t i = 0; i < 16; i++) {
+            tmp[i] = remap[*(c + i)];
+        }
+        m128 str_c = loadu128(tmp);
+
+        u32 z = movemask128(eq128(str_w, str_c));
+        pos = ctz32(~z);
+        assert(pos <= 16);
+
+        if (pos < 16) {
+            goto normal;
+        }
+
+        sym += 16;
+        c += 16;
+        len_w -= 16;
+        len_c -= 16;
+    }
+
+    pos = 0;
+    // at least one in (0, 16).
+    u32 loadLength_w = MIN(len_w, 16);
+    u32 loadLength_c = MIN(len_c, 16);
+    m128 str_w = loadbytes128(sym, loadLength_w);
+    for (size_t i = 0; i < loadLength_c; i++) {
+        tmp[i] = remap[*(c + i)];
+    }
+    m128 str_c = loadbytes128(tmp, loadLength_c);
+
+    u32 z = movemask128(eq128(str_w, str_c));
+    pos = ctz32(~z);
+
+    pos = MIN(pos, MIN(loadLength_w, loadLength_c));
+
+    if (loadLength_w <= loadLength_c) {
+        assert(pos <= loadLength_w);
+        // successful matching.
+        if (pos == loadLength_w) {
+            c -= 1;
+            successful = 1;
+        }
+        // failure, do nothing.
+    } else {
+        assert(pos <= loadLength_c);
+        // successful partial matching.
+        if (pos == loadLength_c) {
+            c -= 1;
+            goto partial;
+        }
+        // failure, do nothing.
+    }
+
+normal:
+    *offset = 0;
+    if (qstate != NULL) {
+        // Internal relative offset.
+        unaligned_store_u16(qstate + 2, *offset);
+    }
+    c += pos;
+    *c_inout = c;
+    return successful ? *trans : *(trans + 1 + remap[*c]);
+
+partial:
+    *offset = sym - symbols + pos;
+    if (qstate != NULL) {
+        // Internal relative offset.
+        unaligned_store_u16(qstate + 2, *offset);
+    }
+    c += pos;
+    *c_inout = c;
+    return *s;
+}

src/nfa/mcclellan_internal.h

@@ -50,6 +50,16 @@ extern "C"
 #define SHERMAN_CHARS_OFFSET           4
 #define SHERMAN_STATES_OFFSET(sso_len) (4 + (sso_len))
 
+#define WIDE_STATE 2
+#define WIDE_ENTRY_OFFSET8(weo_pos) (2 + (weo_pos))
+#define WIDE_ENTRY_OFFSET16(weo_pos) (3 + (weo_pos))
+
+#define WIDE_WIDTH_OFFSET 0
+#define WIDE_SYMBOL_OFFSET8 1
+#define WIDE_TRANSITION_OFFSET8(wto_width) (1 + (wto_width))
+#define WIDE_SYMBOL_OFFSET16 2
+#define WIDE_TRANSITION_OFFSET16(wto_width) (2 + (wto_width))
+
 struct report_list {
     u32 count;
     ReportID report[];
@@ -79,13 +89,17 @@ struct mcclellan {
     u16 accel_limit_8; /**< 8 bit, lowest accelerable state */
     u16 accept_limit_8; /**< 8 bit, lowest accept state */
     u16 sherman_limit; /**< lowest sherman state */
+    u16 wide_limit; /**< 8/16 bit, lowest wide head state */
     u8  alphaShift;
     u8  flags;
     u8  has_accel; /**< 1 iff there are any accel plans */
+    u8  has_wide; /**< 1 iff there exists any wide state */
     u8  remap[256]; /**< remaps characters to a smaller alphabet */
     ReportID arb_report; /**< one of the accepts that this dfa may raise */
     u32 accel_offset; /**< offset of accel structures from start of McClellan */
     u32 haig_offset; /**< reserved for use by Haig, relative to start of NFA */
+    u32 wide_offset; /**< offset of the wide state entries to the start of the
+                      * nfa structure */
 };
 
 static really_inline
@@ -106,6 +120,43 @@ char *findMutableShermanState(char *sherman_base_offset, u16 sherman_base,
     return sherman_base_offset + SHERMAN_FIXED_SIZE * (s - sherman_base);
 }
 
+static really_inline
+const char *findWideEntry8(UNUSED const struct mcclellan *m,
+                           const char *wide_base, u32 wide_limit, u32 s) {
+    UNUSED u8 type = *(const u8 *)wide_base;
+    assert(type == WIDE_STATE);
+    const u32 entry_offset
+        = *(const u32 *)(wide_base
+        + WIDE_ENTRY_OFFSET8((s - wide_limit) * sizeof(u32)));
+
+    const char *rv = wide_base + entry_offset;
+    assert(rv < (const char *)m + m->length - sizeof(struct NFA));
+    return rv;
+}
+
+static really_inline
+const char *findWideEntry16(UNUSED const struct mcclellan *m,
+                            const char *wide_base, u32 wide_limit, u32 s) {
+    UNUSED u8 type = *(const u8 *)wide_base;
+    assert(type == WIDE_STATE);
+    const u32 entry_offset
+        = *(const u32 *)(wide_base
+        + WIDE_ENTRY_OFFSET16((s - wide_limit) * sizeof(u32)));
+
+    const char *rv = wide_base + entry_offset;
+    assert(rv < (const char *)m + m->length - sizeof(struct NFA));
+    return rv;
+}
+
+static really_inline
+char *findMutableWideEntry16(char *wide_base, u32 wide_limit, u32 s) {
+    u32 entry_offset
+        = *(const u32 *)(wide_base
+        + WIDE_ENTRY_OFFSET16((s - wide_limit) * sizeof(u32)));
+
+    return wide_base + entry_offset;
+}
+
 #ifdef __cplusplus
 }
 #endif

src/nfa/mcclellancompile.cpp

@@ -56,13 +56,19 @@
 #include <cstring>
 #include <map>
 #include <memory>
+#include <queue>
 #include <set>
 #include <vector>
 
 #include <boost/range/adaptor/map.hpp>
 
+#include "mcclellandump.h"
+#include "util/dump_util.h"
+#include "util/dump_charclass.h"
+
 using namespace std;
 using boost::adaptors::map_keys;
+using boost::dynamic_bitset;
 
 #define ACCEL_DFA_MAX_OFFSET_DEPTH 4
 
@@ -82,6 +88,8 @@ namespace /* anon */ {
 struct dstate_extra {
     u16 daddytaken = 0;
     bool shermanState = false;
+    bool wideState = false;
+    bool wideHead = false;
 };
 
 struct dfa_info {
@@ -89,6 +97,8 @@ struct dfa_info {
     raw_dfa &raw;
     vector<dstate> &states;
     vector<dstate_extra> extra;
+    vector<vector<dstate_id_t>> wide_state_chain;
+    vector<vector<symbol_t>> wide_symbol_chain;
     const u16 alpha_size; /* including special symbols */
     const array<u16, ALPHABET_SIZE> &alpha_remap;
     const u16 impl_alpha_size;
@@ -112,6 +122,14 @@ struct dfa_info {
         return extra[raw_id].shermanState;
     }
 
+    bool is_widestate(dstate_id_t raw_id) const {
+        return extra[raw_id].wideState;
+    }
+
+    bool is_widehead(dstate_id_t raw_id) const {
+        return extra[raw_id].wideHead;
+    }
+
     size_t size(void) const { return states.size(); }
 };
 
@@ -124,6 +142,35 @@ u8 dfa_info::getAlphaShift() const {
     }
 }
 
+struct state_prev_info {
+    vector<vector<dstate_id_t>> prev_vec;
+    explicit state_prev_info(size_t alpha_size) : prev_vec(alpha_size) {}
+};
+
+struct DfaPrevInfo {
+    u16 impl_alpha_size;
+    u16 state_num;
+    vector<state_prev_info> states;
+    set<dstate_id_t> accepts;
+
+    explicit DfaPrevInfo(raw_dfa &rdfa);
+};
+
+DfaPrevInfo::DfaPrevInfo(raw_dfa &rdfa)
+    : impl_alpha_size(rdfa.getImplAlphaSize()), state_num(rdfa.states.size()),
+      states(state_num, state_prev_info(impl_alpha_size)){
+    for (size_t i = 0; i < states.size(); i++) {
+        for (symbol_t sym = 0; sym < impl_alpha_size; sym++) {
+            dstate_id_t curr = rdfa.states[i].next[sym];
+            states[curr].prev_vec[sym].push_back(i);
+        }
+        if (!rdfa.states[i].reports.empty()
+            || !rdfa.states[i].reports_eod.empty()) {
+            DEBUG_PRINTF("accept raw state: %ld\n", i);
+            accepts.insert(i);
+        }
+    }
+}
 } // namespace
 
 static
@@ -151,6 +198,11 @@ void markEdges(NFA *n, u16 *succ_table, const dfa_info &info) {
         for (size_t j = 0; j < alphaSize; j++) {
             size_t c_prime = (i << alphaShift) + j;
 
+            // wide state has no aux structure.
+            if (m->has_wide && succ_table[c_prime] >= m->wide_limit) {
+                continue;
+            }
+
             mstate_aux *aux = getAux(n, succ_table[c_prime]);
 
             if (aux->accept) {
@@ -165,7 +217,8 @@ void markEdges(NFA *n, u16 *succ_table, const dfa_info &info) {
 
     /* handle the sherman states */
     char *sherman_base_offset = (char *)n + m->sherman_offset;
-    for (u16 j = m->sherman_limit; j < m->state_count; j++) {
+    u16 sherman_ceil = m->has_wide == 1 ? m->wide_limit : m->state_count;
+    for (u16 j = m->sherman_limit; j < sherman_ceil; j++) {
         char *sherman_cur
             = findMutableShermanState(sherman_base_offset, m->sherman_limit, j);
         assert(*(sherman_cur + SHERMAN_TYPE_OFFSET) == SHERMAN_STATE);
@@ -174,6 +227,11 @@ void markEdges(NFA *n, u16 *succ_table, const dfa_info &info) {
 
         for (u8 i = 0; i < len; i++) {
             u16 succ_i = unaligned_load_u16((u8 *)&succs[i]);
+            // wide state has no aux structure.
+            if (m->has_wide && succ_i >= m->wide_limit) {
+                continue;
+            }
+
             mstate_aux *aux = getAux(n, succ_i);
 
             if (aux->accept) {
@@ -187,6 +245,49 @@ void markEdges(NFA *n, u16 *succ_table, const dfa_info &info) {
             unaligned_store_u16((u8 *)&succs[i], succ_i);
         }
     }
+
+    /* handle the wide states */
+    if (m->has_wide) {
+        u32 wide_limit = m->wide_limit;
+        char *wide_base = (char *)n + m->wide_offset;
+        assert(*wide_base == WIDE_STATE);
+        u16 wide_number = verify_u16(info.wide_symbol_chain.size());
+        // traverse over wide head states.
+        for (u16 j = wide_limit; j < wide_limit + wide_number; j++) {
+            char *wide_cur
+                = findMutableWideEntry16(wide_base, wide_limit, j);
+            u16 width = *(const u16 *)(wide_cur + WIDE_WIDTH_OFFSET);
+            u16 *trans = (u16 *)(wide_cur + WIDE_TRANSITION_OFFSET16(width));
+
+            // check successful transition
+            u16 next = unaligned_load_u16((u8 *)trans);
+            if (next >= wide_limit) {
+                continue;
+            }
+            mstate_aux *aux = getAux(n, next);
+            if (aux->accept) {
+                next |= ACCEPT_FLAG;
+            }
+            if (aux->accel_offset) {
+                next |= ACCEL_FLAG;
+            }
+            unaligned_store_u16((u8 *)trans, next);
+            trans ++;
+
+            // check failure transition
+            for (symbol_t k = 0; k < alphaSize; k++) {
+                u16 next_k = unaligned_load_u16((u8 *)&trans[k]);
+                mstate_aux *aux_k = getAux(n, next_k);
+                if (aux_k->accept) {
+                    next_k |= ACCEPT_FLAG;
+                }
+                if (aux_k->accel_offset) {
+                    next_k |= ACCEL_FLAG;
+                }
+                unaligned_store_u16((u8 *)&trans[k], next_k);
+            }
+        }
+    }
 }
 
 u32 mcclellan_build_strat::max_allowed_offset_accel() const {
@@ -232,6 +333,19 @@ void populateBasicInfo(size_t state_size, const dfa_info &info,
     m->start_anchored = info.implId(info.raw.start_anchored);
     m->start_floating = info.implId(info.raw.start_floating);
     m->has_accel = accel_count ? 1 : 0;
+    m->has_wide = info.wide_state_chain.size() > 0 ? 1 : 0;
+
+    if (state_size == sizeof(u8) && m->has_wide == 1) {
+        // allocate 1 more byte for wide state use.
+        nfa->scratchStateSize += sizeof(u8);
+        nfa->streamStateSize += sizeof(u8);
+    }
+
+    if (state_size == sizeof(u16) && m->has_wide == 1) {
+        // allocate 2 more bytes for wide state use.
+        nfa->scratchStateSize += sizeof(u16);
+        nfa->streamStateSize += sizeof(u16);
+    }
 
     if (single) {
         m->flags |= MCCLELLAN_FLAG_SINGLE;
@@ -404,6 +518,23 @@ size_t calcShermanRegionSize(const dfa_info &info) {
     return ROUNDUP_16(rv);
 }
 
+static
+size_t calcWideRegionSize(const dfa_info &info) {
+    if (info.wide_state_chain.empty()) {
+        return 0;
+    }
+
+    // wide info header
+    size_t rv = info.wide_symbol_chain.size() * sizeof(u32) + 3;
+
+    // wide info body
+    for (const auto &chain : info.wide_symbol_chain) {
+        rv += chain.size() + (info.impl_alpha_size + 1) * sizeof(u16) + 2;
+    }
+
+    return ROUNDUP_16(rv);
+}
+
 static
 void fillInAux(mstate_aux *aux, dstate_id_t i, const dfa_info &info,
                const vector<u32> &reports, const vector<u32> &reports_eod,
@@ -418,42 +549,60 @@ void fillInAux(mstate_aux *aux, dstate_id_t i, const dfa_info &info,
 
 /* returns false on error */
 static
-bool allocateFSN16(dfa_info &info, dstate_id_t *sherman_base) {
+bool allocateFSN16(dfa_info &info, dstate_id_t *sherman_base,
+                   dstate_id_t *wide_limit) {
     info.states[0].impl_id = 0; /* dead is always 0 */
 
     vector<dstate_id_t> norm;
     vector<dstate_id_t> sherm;
+    vector<dstate_id_t> wideHead;
+    vector<dstate_id_t> wideState;
 
     if (info.size() > (1 << 16)) {
         DEBUG_PRINTF("too many states\n");
-        *sherman_base = 0;
+        *wide_limit = 0;
         return false;
     }
 
     for (u32 i = 1; i < info.size(); i++) {
-        if (info.is_sherman(i)) {
+        if (info.is_widehead(i)) {
+            wideHead.push_back(i);
+        } else if (info.is_widestate(i)) {
+            wideState.push_back(i);
+        } else if (info.is_sherman(i)) {
             sherm.push_back(i);
         } else {
             norm.push_back(i);
         }
     }
 
-    dstate_id_t next_norm = 1;
+    dstate_id_t next = 1;
     for (const dstate_id_t &s : norm) {
-        info.states[s].impl_id = next_norm++;
+        DEBUG_PRINTF("[norm] mapping state %u to %u\n", s, next);
+        info.states[s].impl_id = next++;
     }
 
-    *sherman_base = next_norm;
+    *sherman_base = next;
-    dstate_id_t next_sherman = next_norm;
-
     for (const dstate_id_t &s : sherm) {
-        info.states[s].impl_id = next_sherman++;
+        DEBUG_PRINTF("[sherm] mapping state %u to %u\n", s, next);
+        info.states[s].impl_id = next++;
+    }
+
+    *wide_limit = next;
+    for (const dstate_id_t &s : wideHead) {
+        DEBUG_PRINTF("[widehead] mapping state %u to %u\n", s, next);
+        info.states[s].impl_id = next++;
+    }
+
+    for (const dstate_id_t &s : wideState) {
+        DEBUG_PRINTF("[wide] mapping state %u to %u\n", s, next);
+        info.states[s].impl_id = next++;
     }
 
     /* Check to see if we haven't over allocated our states */
-    DEBUG_PRINTF("next sherman %u masked %u\n", next_sherman,
+    DEBUG_PRINTF("next sherman %u masked %u\n", next,
-                 (dstate_id_t)(next_sherman & STATE_MASK));
+                 (dstate_id_t)(next & STATE_MASK));
-    return (next_sherman - 1) == ((next_sherman - 1) & STATE_MASK);
+    return (next - 1) == ((next - 1) & STATE_MASK);
 }
 
 static
@@ -470,12 +619,16 @@ bytecode_ptr<NFA> mcclellanCompile16(dfa_info &info, const CompileContext &cc,
     assert(alphaShift <= 8);
 
     u16 count_real_states;
-    if (!allocateFSN16(info, &count_real_states)) {
+    u16 wide_limit;
+    if (!allocateFSN16(info, &count_real_states, &wide_limit)) {
         DEBUG_PRINTF("failed to allocate state numbers, %zu states total\n",
                      info.size());
         return nullptr;
     }
 
+    DEBUG_PRINTF("count_real_states: %d\n", count_real_states);
+    DEBUG_PRINTF("non_wide_states: %d\n", wide_limit);
+
     auto ri = info.strat.gatherReports(reports, reports_eod, &single, &arb);
     map<dstate_id_t, AccelScheme> accel_escape_info
             = info.strat.getAccelInfo(cc.grey);
@@ -483,7 +636,7 @@ bytecode_ptr<NFA> mcclellanCompile16(dfa_info &info, const CompileContext &cc,
     size_t tran_size = (1 << info.getAlphaShift())
         * sizeof(u16) * count_real_states;
 
-    size_t aux_size = sizeof(mstate_aux) * info.size();
+    size_t aux_size = sizeof(mstate_aux) * wide_limit;
 
     size_t aux_offset = ROUNDUP_16(sizeof(NFA) + sizeof(mcclellan) + tran_size);
     size_t accel_size = info.strat.accelSize() * accel_escape_info.size();
@@ -491,12 +644,24 @@ bytecode_ptr<NFA> mcclellanCompile16(dfa_info &info, const CompileContext &cc,
                                     + ri->getReportListSize(), 32);
     size_t sherman_offset = ROUNDUP_16(accel_offset + accel_size);
     size_t sherman_size = calcShermanRegionSize(info);
-
+    size_t wide_offset = ROUNDUP_16(sherman_offset + sherman_size);
-    size_t total_size = sherman_offset + sherman_size;
+    size_t wide_size = calcWideRegionSize(info);
+    size_t total_size = wide_offset + wide_size;
 
     accel_offset -= sizeof(NFA); /* adj accel offset to be relative to m */
     assert(ISALIGNED_N(accel_offset, alignof(union AccelAux)));
 
+    DEBUG_PRINTF("aux_offset %zu\n", aux_offset);
+    DEBUG_PRINTF("aux_size %zu\n", aux_size);
+    DEBUG_PRINTF("rl size %u\n", ri->getReportListSize());
+    DEBUG_PRINTF("accel_offset %zu\n", accel_offset + sizeof(NFA));
+    DEBUG_PRINTF("accel_size %zu\n", accel_size);
+    DEBUG_PRINTF("sherman_offset %zu\n", sherman_offset);
+    DEBUG_PRINTF("sherman_size %zu\n", sherman_size);
+    DEBUG_PRINTF("wide_offset %zu\n", wide_offset);
+    DEBUG_PRINTF("wide_size %zu\n", wide_size);
+    DEBUG_PRINTF("total_size %zu\n", total_size);
+
     auto nfa = make_zeroed_bytecode_ptr<NFA>(total_size);
     char *nfa_base = (char *)nfa.get();
 
@@ -511,6 +676,9 @@ bytecode_ptr<NFA> mcclellanCompile16(dfa_info &info, const CompileContext &cc,
     mstate_aux *aux = (mstate_aux *)(nfa_base + aux_offset);
     mcclellan *m = (mcclellan *)getMutableImplNfa(nfa.get());
 
+    m->wide_limit = wide_limit;
+    m->wide_offset = wide_offset;
+
     /* copy in the mc header information */
     m->sherman_offset = sherman_offset;
     m->sherman_end = total_size;
@@ -518,7 +686,7 @@ bytecode_ptr<NFA> mcclellanCompile16(dfa_info &info, const CompileContext &cc,
 
     /* do normal states */
     for (size_t i = 0; i < info.size(); i++) {
-        if (info.is_sherman(i)) {
+        if (info.is_sherman(i) || info.is_widestate(i)) {
             continue;
         }
 
@@ -556,6 +724,7 @@ bytecode_ptr<NFA> mcclellanCompile16(dfa_info &info, const CompileContext &cc,
         mstate_aux *this_aux = getAux(nfa.get(), fs);
 
         assert(fs >= count_real_states);
+        assert(fs < wide_limit);
 
         char *curr_sherman_entry
             = sherman_table + (fs - m->sherman_limit) * SHERMAN_FIXED_SIZE;
@@ -599,6 +768,70 @@ bytecode_ptr<NFA> mcclellanCompile16(dfa_info &info, const CompileContext &cc,
         }
     }
 
+    if (!info.wide_state_chain.empty()) {
+        /* do wide states using info */
+        u16 wide_number = verify_u16(info.wide_symbol_chain.size());
+        char *wide_base = nfa_base + m->wide_offset;
+        assert(ISALIGNED_16(wide_base));
+
+        char *wide_top = wide_base;
+        *(u8 *)(wide_top++) = WIDE_STATE;
+        *(u16 *)(wide_top) = wide_number;
+        wide_top += 2;
+
+        char *curr_wide_entry = wide_top + wide_number * sizeof(u32);
+        u32 *wide_offset_list = (u32 *)wide_top;
+
+        /* get the order of writing wide states */
+        vector<size_t> order(wide_number);
+        for (size_t i = 0; i < wide_number; i++) {
+            dstate_id_t head = info.wide_state_chain[i].front();
+            size_t pos = info.implId(head) - m->wide_limit;
+            order[pos] = i;
+        }
+
+        for (size_t i : order) {
+            vector<dstate_id_t> &state_chain = info.wide_state_chain[i];
+            vector<symbol_t> &symbol_chain = info.wide_symbol_chain[i];
+
+            u16 width = verify_u16(symbol_chain.size());
+            *(u16 *)(curr_wide_entry + WIDE_WIDTH_OFFSET) = width;
+            u8 *chars = (u8 *)(curr_wide_entry + WIDE_SYMBOL_OFFSET16);
+
+            // store wide state symbol chain
+            for (size_t j = 0; j < width; j++) {
+                *(chars++) = verify_u8(symbol_chain[j]);
+            }
+
+            // store wide state transition table
+            u16 *trans = (u16 *)(curr_wide_entry
+                                + WIDE_TRANSITION_OFFSET16(width));
+            dstate_id_t tail = state_chain[width - 1];
+            symbol_t last = symbol_chain[width -1];
+            dstate_id_t tran = info.states[tail].next[last];
+            // 1. successful transition
+            *trans++ = info.implId(tran);
+            // 2. failure transition
+            for (size_t j = 0; verify_u16(j) < width - 1; j++) {
+                if (symbol_chain[j] != last) {
+                    tran = info.states[state_chain[j]].next[last];
+                }
+            }
+            for (symbol_t sym = 0; sym < info.impl_alpha_size; sym++) {
+                if (sym != last) {
+                    *trans++ = info.implId(info.states[tail].next[sym]);
+                }
+                else {
+                    *trans++ = info.implId(tran);
+                }
+            }
+
+            *wide_offset_list++ = verify_u32(curr_wide_entry - wide_base);
+
+            curr_wide_entry = (char *)trans;
+        }
+    }
+
     markEdges(nfa.get(), succ_table, info);
 
     if (accel_states && nfa) {
@@ -844,12 +1077,16 @@ void find_better_daddy(dfa_info &info, dstate_id_t curr_id, bool using8bit,
     if (trust_daddy_states) {
         // Use the daddy already set for this state so long as it isn't already
         // a Sherman state.
-        if (!info.is_sherman(currState.daddy)) {
+        dstate_id_t daddy = currState.daddy;
+        if (!info.is_sherman(daddy) && !info.is_widestate(daddy)) {
             hinted.insert(currState.daddy);
         } else {
             // Fall back to granddaddy, which has already been processed (due
             // to BFS ordering) and cannot be a Sherman state.
             dstate_id_t granddaddy = info.states[currState.daddy].daddy;
+            if (info.is_widestate(granddaddy)) {
+                return;
+            }
             assert(!info.is_sherman(granddaddy));
             hinted.insert(granddaddy);
         }
@@ -861,7 +1098,7 @@ void find_better_daddy(dfa_info &info, dstate_id_t curr_id, bool using8bit,
         assert(donor < curr_id);
         u32 score = 0;
 
-        if (info.is_sherman(donor)) {
+        if (info.is_sherman(donor) || info.is_widestate(donor)) {
             continue;
         }
 
@@ -934,6 +1171,290 @@ bool is_cyclic_near(const raw_dfa &raw, dstate_id_t root) {
     return false;
 }
 
+/* \brief Test for only-one-predecessor property. */
+static
+bool check_property1(const DfaPrevInfo &info, const u16 impl_alpha_size,
+                     const dstate_id_t curr_id, dstate_id_t &prev_id,
+                     symbol_t &prev_sym) {
+    u32 num_prev = 0;
+    bool test_p1 = false;
+
+    for (symbol_t sym = 0; sym < impl_alpha_size; sym++) {
+        num_prev += info.states[curr_id].prev_vec[sym].size();
+        DEBUG_PRINTF("Check symbol: %u, with its vector size: %lu\n", sym,
+                     info.states[curr_id].prev_vec[sym].size());
+        if (num_prev == 1 && !test_p1) {
+            test_p1 = true;
+            prev_id = info.states[curr_id].prev_vec[sym].front(); //[0] for sure???
+            prev_sym = sym;
+        }
+    }
+
+    return num_prev == 1;
+}
+
+/* \brief Test for same-failure-action property. */
+static
+bool check_property2(const raw_dfa &rdfa, const u16 impl_alpha_size,
+                     const dstate_id_t curr_id, const dstate_id_t prev_id,
+                     const symbol_t curr_sym, const symbol_t prev_sym) {
+    const dstate &prevState = rdfa.states[prev_id];
+    const dstate &currState = rdfa.states[curr_id];
+
+    // Compare transition tables between currState and prevState.
+    u16 score = 0;
+    for (symbol_t sym = 0; sym < impl_alpha_size; sym++) {
+        if (currState.next[sym] == prevState.next[sym]
+            && sym != curr_sym && sym != prev_sym) {
+            score++;
+        }
+    }
+    DEBUG_PRINTF("(Score: %u/%u)\n", score, impl_alpha_size);
+
+    // 2 cases.
+    if (curr_sym != prev_sym && score >= impl_alpha_size - 2
+        && currState.next[prev_sym] == prevState.next[curr_sym]) {
+        return true;
+    } else if (curr_sym == prev_sym && score == impl_alpha_size - 1) {
+        return true;
+    }
+    return false;
+}
+
+/* \brief Check whether adding current prev_id will generate a circle.*/
+static
+bool check_circle(const DfaPrevInfo &info, const u16 impl_alpha_size,
+                  const vector<dstate_id_t> &chain, const dstate_id_t id) {
+    const vector<vector<dstate_id_t>> &prev_vec = info.states[id].prev_vec;
+    const dstate_id_t tail = chain.front();
+    for (symbol_t sym = 0; sym < impl_alpha_size; sym++) {
+        auto iter = find(prev_vec[sym].begin(), prev_vec[sym].end(), tail);
+        if (iter != prev_vec[sym].end()) {
+            // Tail is one of id's predecessors, forming a circle.
+            return true;
+        }
+    }
+    return false;
+}
+
+/* \brief Returns a chain of state ids and symbols. */
+static
+dstate_id_t find_chain_candidate(const raw_dfa &rdfa, const DfaPrevInfo &info,
+                                 const dstate_id_t curr_id,
+                                 const symbol_t curr_sym,
+                                 vector<dstate_id_t> &temp_chain) {
+    //Record current id first.
+    temp_chain.push_back(curr_id);
+
+    const u16 size = info.impl_alpha_size;
+
+    // Stop when entering root cloud.
+    if (rdfa.start_anchored != DEAD_STATE
+        && is_cyclic_near(rdfa, rdfa.start_anchored)
+        && curr_id < size) {
+       return curr_id;
+    }
+    if (rdfa.start_floating != DEAD_STATE
+        && curr_id >= rdfa.start_floating
+        && curr_id < rdfa.start_floating + size * 3) {
+        return curr_id;
+    }
+
+    // Stop when reaching anchored or floating.
+    if (curr_id == rdfa.start_anchored || curr_id == rdfa.start_floating) {
+        return curr_id;
+    }
+
+    dstate_id_t prev_id = 0;
+    symbol_t prev_sym = ALPHABET_SIZE;
+
+    // Check the only-one-predecessor property.
+    if (!check_property1(info, size, curr_id, prev_id, prev_sym)) {
+        return curr_id;
+    }
+    assert(prev_id != 0 && prev_sym != ALPHABET_SIZE);
+    DEBUG_PRINTF("(P1 test passed.)\n");
+
+    // Circle testing for the prev_id that passes the P1 test.
+    if (check_circle(info, size, temp_chain, prev_id)) {
+        DEBUG_PRINTF("(A circle is found.)\n");
+        return curr_id;
+    }
+
+    // Check the same-failure-action property.
+    if (!check_property2(rdfa, size, curr_id, prev_id, curr_sym, prev_sym)) {
+        return curr_id;
+    }
+    DEBUG_PRINTF("(P2 test passed.)\n");
+
+    if (!rdfa.states[prev_id].reports.empty()
+        || !rdfa.states[prev_id].reports_eod.empty()) {
+        return curr_id;
+    } else {
+        return find_chain_candidate(rdfa, info, prev_id, prev_sym, temp_chain);
+    }
+}
+
+/* \brief Always store the non-extensible chains found till now. */
+static
+bool store_chain_longest(vector<vector<dstate_id_t>> &candidate_chain,
+                         vector<dstate_id_t> &temp_chain,
+                         dynamic_bitset<> &added, bool head_is_new) {
+    dstate_id_t head = temp_chain.front();
+    u16 length = temp_chain.size();
+
+    if (head_is_new) {
+        DEBUG_PRINTF("This is a new chain!\n");
+
+        // Add this new chain and get it marked.
+        candidate_chain.push_back(temp_chain);
+
+        for (auto &id : temp_chain) {
+            DEBUG_PRINTF("(Marking s%u ...)\n", id);
+            added.set(id);
+        }
+
+        return true;
+    }
+
+    DEBUG_PRINTF("This is a longer chain!\n");
+    assert(!candidate_chain.empty());
+
+    auto chain = find_if(candidate_chain.begin(), candidate_chain.end(),
+                         [&](const vector<dstate_id_t> &it) {
+                            return it.front() == head;
+                         });
+
+    // Not a valid head, just do nothing and return.
+    if (chain == candidate_chain.end()) {
+        return false;
+    }
+
+    u16 len = chain->size();
+
+    if (length > len) {
+        // Find out the branch node first.
+        size_t piv = 0;
+        for (; piv < length; piv++) {
+            if ((*chain)[piv] != temp_chain[piv]) {
+                break;
+            }
+        }
+
+        for (size_t j = piv + 1; j < length; j++) {
+            DEBUG_PRINTF("(Marking s%u (new branch) ...)\n", temp_chain[j]);
+            added.set(temp_chain[j]);
+        }
+
+        // Unmark old unuseful nodes.
+        // (Except the tail node, which is in working queue)
+        for (size_t j = piv + 1; j < verify_u16(len - 1); j++) {
+            DEBUG_PRINTF("(UnMarking s%u (old branch)...)\n", (*chain)[j]);
+            added.reset((*chain)[j]);
+        }
+
+        chain->assign(temp_chain.begin(), temp_chain.end());
+    }
+
+    return false;
+}
+
+/* \brief Generate wide_symbol_chain from wide_state_chain. */
+static
+void generate_symbol_chain(dfa_info &info, vector<symbol_t> &chain_tail) {
+    raw_dfa &rdfa = info.raw;
+    assert(chain_tail.size() == info.wide_state_chain.size());
+
+    for (size_t i = 0; i < info.wide_state_chain.size(); i++) {
+        vector<dstate_id_t> &state_chain = info.wide_state_chain[i];
+        vector<symbol_t> symbol_chain;
+
+        info.extra[state_chain[0]].wideHead = true;
+        size_t width = state_chain.size() - 1;
+
+        for (size_t j = 0; j < width; j++) {
+            dstate_id_t curr_id = state_chain[j];
+            dstate_id_t next_id = state_chain[j + 1];
+
+            // The last state of the chain doesn't belong to a wide state.
+            info.extra[curr_id].wideState = true;
+
+            // The tail symbol comes from vector chain_tail;
+            if (j == width - 1) {
+                symbol_chain.push_back(chain_tail[i]);
+            } else {
+                for (symbol_t sym = 0; sym < info.impl_alpha_size; sym++) {
+                    if (rdfa.states[curr_id].next[sym] == next_id) {
+                        symbol_chain.push_back(sym);
+                        break;
+                    }
+                }
+            }
+        }
+
+        info.wide_symbol_chain.push_back(symbol_chain);
+    }
+}
+
+/* \brief Find potential regions of states to be packed into wide states. */
+static
+void find_wide_state(dfa_info &info) {
+    DfaPrevInfo dinfo(info.raw);
+    queue<dstate_id_t> work_queue;
+
+    dynamic_bitset<> added(info.raw.states.size());
+    for (auto it : dinfo.accepts) {
+        work_queue.push(it);
+        added.set(it);
+    }
+
+    vector<symbol_t> chain_tail;
+    while (!work_queue.empty()) {
+        dstate_id_t curr_id = work_queue.front();
+        work_queue.pop();
+        DEBUG_PRINTF("Newly popped state: s%u\n", curr_id);
+
+        for (symbol_t sym = 0; sym < dinfo.impl_alpha_size; sym++) {
+            for (auto info_it : dinfo.states[curr_id].prev_vec[sym]) {
+                if (added.test(info_it)) {
+                    DEBUG_PRINTF("(s%u already marked.)\n", info_it);
+                    continue;
+                }
+
+                vector<dstate_id_t> temp_chain;
+                // Head is a state failing the test of the chain.
+                dstate_id_t head = find_chain_candidate(info.raw, dinfo,
+                                                        info_it, sym,
+                                                        temp_chain);
+
+                // A candidate chain should contain 8 substates at least.
+                if (temp_chain.size() < 8) {
+                    DEBUG_PRINTF("(Not enough substates, continue.)\n");
+                    continue;
+                }
+
+                bool head_is_new = !added.test(head);
+                if (head_is_new) {
+                    added.set(head);
+                    work_queue.push(head);
+                    DEBUG_PRINTF("Newly pushed state: s%u\n", head);
+                }
+
+                reverse(temp_chain.begin(), temp_chain.end());
+                temp_chain.push_back(curr_id);
+
+                assert(head > 0 && head == temp_chain.front());
+                if (store_chain_longest(info.wide_state_chain, temp_chain,
+                                        added, head_is_new)) {
+                    chain_tail.push_back(sym);
+                }
+            }
+        }
+    }
+
+    generate_symbol_chain(info, chain_tail);
+}
+
 bytecode_ptr<NFA> mcclellanCompile_i(raw_dfa &raw, accel_dfa_build_strat &strat,
                                      const CompileContext &cc,
                                      bool trust_daddy_states,
@@ -952,11 +1473,19 @@ bytecode_ptr<NFA> mcclellanCompile_i(raw_dfa &raw, accel_dfa_build_strat &strat,
 
     bytecode_ptr<NFA> nfa;
     if (!using8bit) {
+        if (cc.grey.allowWideStates && strat.getType() == McClellan
+            && !is_triggered(raw.kind)) {
+            find_wide_state(info);
+        }
+
         u16 total_daddy = 0;
         bool any_cyclic_near_anchored_state
             = is_cyclic_near(raw, raw.start_anchored);
 
         for (u32 i = 0; i < info.size(); i++) {
+            if (info.is_widestate(i)) {
+                continue;
+            }
             find_better_daddy(info, i, using8bit,
                               any_cyclic_near_anchored_state,
                               trust_daddy_states, cc.grey);

src/nfa/mcclellancompile.h

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2015-2017, Intel Corporation
+ * Copyright (c) 2015-2018, Intel Corporation
  *
  * Redistribution and use in source and binary forms, with or without
  * modification, are permitted provided that the following conditions are met:
@@ -60,6 +60,7 @@ public:
     u32 max_allowed_offset_accel() const override;
     u32 max_stop_char() const override;
     u32 max_floating_stop_char() const override;
+    DfaType getType() const override { return McClellan; }
 
 private:
     raw_dfa &rdfa;

src/nfa/shengcompile.h

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2017, Intel Corporation
+ * Copyright (c) 2016-2018, Intel Corporation
  *
  * Redistribution and use in source and binary forms, with or without
  * modification, are permitted provided that the following conditions are met:
@@ -61,6 +61,7 @@ public:
     u32 max_allowed_offset_accel() const override;
     u32 max_stop_char() const override;
     u32 max_floating_stop_char() const override;
+    DfaType getType() const override { return Sheng; }
 
 private:
     raw_dfa &rdfa;