From 805a550a0aa5b137eb989558d7969cc2964954d6 Mon Sep 17 00:00:00 2001
From: "Hong, Yang A" <yang.a.hong@intel.com>
Date: Mon, 24 Dec 2018 04:49:47 +0800
Subject: [PATCH] mcclellan: wide state fixes for sanitisers and accept state
 construction

---
 src/nfa/mcclellan.c             | 10 +++++-----
 src/nfa/mcclellan_common_impl.h |  2 +-
 src/nfa/mcclellan_internal.h    |  4 ++--
 src/nfa/mcclellancompile.cpp    | 30 +++++++++++++++++-------------
 4 files changed, 25 insertions(+), 21 deletions(-)

diff --git a/src/nfa/mcclellan.c b/src/nfa/mcclellan.c
index 1521de5b..71f71e32 100644
--- a/src/nfa/mcclellan.c
+++ b/src/nfa/mcclellan.c
@@ -1197,7 +1197,7 @@ char nfaExecMcClellan16_initCompressedState(const struct NFA *nfa, u64a offset,
 
     // new byte
     if (m->has_wide) {
-        *((u16 *)state + 1) = 0;
+        unaligned_store_u16((u16 *)state + 1, 0);
     }
 
     if (s) {
@@ -1236,7 +1236,7 @@ void nfaExecMcClellan16_SimpStream(const struct NFA *nfa, char *state,
 
         // new byte
         if (m->has_wide) {
-            *((u16 *)state + 1) = 0;
+            unaligned_store_u16((u16 *)state + 1, 0);
         }
     } else {
         s = unaligned_load_u16(state);
@@ -1285,7 +1285,7 @@ char nfaExecMcClellan16_queueInitState(UNUSED const struct NFA *nfa,
 
     // new byte
     if (m->has_wide) {
-        *((u16 *)q->state + 1) = 0;
+        unaligned_store_u16((u16 *)q->state + 1, 0);
     }
     return 0;
 }
@@ -1325,7 +1325,7 @@ char nfaExecMcClellan16_queueCompressState(UNUSED const struct NFA *nfa,
 
     // new byte
     if (m->has_wide) {
-        *((u16 *)dest + 1) = *((const u16 *)src + 1);
+        unaligned_store_u16((u16 *)dest + 1, *((const u16 *)src + 1));
     }
     return 0;
 }
@@ -1344,7 +1344,7 @@ char nfaExecMcClellan16_expandState(UNUSED const struct NFA *nfa, void *dest,
 
     // new byte
     if (m->has_wide) {
-        *((u16 *)dest + 1) = *((const u16 *)src + 1);
+        *((u16 *)dest + 1) =  unaligned_load_u16((const u16 *)src + 1);
     }
     return 0;
 }
diff --git a/src/nfa/mcclellan_common_impl.h b/src/nfa/mcclellan_common_impl.h
index b6af672d..7b0e7f48 100644
--- a/src/nfa/mcclellan_common_impl.h
+++ b/src/nfa/mcclellan_common_impl.h
@@ -88,7 +88,7 @@ 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);
+        *offset = unaligned_load_u16((const u16 *)(qstate + 2));
     }
 
     u8 successful = 0;
diff --git a/src/nfa/mcclellan_internal.h b/src/nfa/mcclellan_internal.h
index 0981f99e..482fdb1b 100644
--- a/src/nfa/mcclellan_internal.h
+++ b/src/nfa/mcclellan_internal.h
@@ -52,13 +52,13 @@ extern "C"
 
 #define WIDE_STATE 2
 #define WIDE_ENTRY_OFFSET8(weo_pos) (2 + (weo_pos))
-#define WIDE_ENTRY_OFFSET16(weo_pos) (3 + (weo_pos))
+#define WIDE_ENTRY_OFFSET16(weo_pos) (4 + (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))
+#define WIDE_TRANSITION_OFFSET16(wto_width) (2 + ROUNDUP_N(wto_width, 2))
 
 struct report_list {
     u32 count;
diff --git a/src/nfa/mcclellancompile.cpp b/src/nfa/mcclellancompile.cpp
index db142f86..c1a4f87f 100644
--- a/src/nfa/mcclellancompile.cpp
+++ b/src/nfa/mcclellancompile.cpp
@@ -261,22 +261,24 @@ void markEdges(NFA *n, u16 *succ_table, const dfa_info &info) {
 
             // check successful transition
             u16 next = unaligned_load_u16((u8 *)trans);
-            if (next >= wide_limit) {
-                continue;
+            if (next < wide_limit) {
+                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);
             }
-            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 ++;
+            trans++;
 
             // check failure transition
             for (symbol_t k = 0; k < alphaSize; k++) {
                 u16 next_k = unaligned_load_u16((u8 *)&trans[k]);
+                if (next_k >= wide_limit) {
+                    continue;
+                }
                 mstate_aux *aux_k = getAux(n, next_k);
                 if (aux_k->accept) {
                     next_k |= ACCEPT_FLAG;
@@ -525,11 +527,12 @@ size_t calcWideRegionSize(const dfa_info &info) {
     }
 
     // wide info header
-    size_t rv = info.wide_symbol_chain.size() * sizeof(u32) + 3;
+    size_t rv = info.wide_symbol_chain.size() * sizeof(u32) + 4;
 
     // wide info body
     for (const auto &chain : info.wide_symbol_chain) {
-        rv += chain.size() + (info.impl_alpha_size + 1) * sizeof(u16) + 2;
+        rv += ROUNDUP_N(chain.size(), 2) +
+              (info.impl_alpha_size + 1) * sizeof(u16) + 2;
     }
 
     return ROUNDUP_16(rv);
@@ -776,6 +779,7 @@ bytecode_ptr<NFA> mcclellanCompile16(dfa_info &info, const CompileContext &cc,
 
         char *wide_top = wide_base;
         *(u8 *)(wide_top++) = WIDE_STATE;
+        wide_top = ROUNDUP_PTR(wide_top, 2);
         *(u16 *)(wide_top) = wide_number;
         wide_top += 2;