Improve scan codegen

src/parser3.h

@@ -285,33 +285,40 @@ struct NumDfa {
   }
   // return value either points to the first byte which does not match, or
   // bufEnd. Leaves the dfa in the last state of the match.
-  const char *scan(const char *buf, const char *bufEnd) {
+#ifdef __x86_64__
+  __attribute__((target_clones("default", "bmi2")))
+#endif
+  const char *
+  scan(const char *buf, const char *bufEnd) {
     auto state_ = state;
     for (;;) {
       constexpr int kStride = 16;
       if (bufEnd - buf < kStride) [[unlikely]] {
         while (buf != bufEnd) {
           uint64_t row = num_dfa_table[uint8_t(*buf)];
-          auto next = row >> (state_ & 63);
-          if ((next & 63) == 0) {
+          auto prev = state_;
+          state_ = (row >> (state_ & 63)) & 63;
+          if (state_ == 0) {
+            state_ = prev;
             break;
           }
-          state_ = next;
           ++buf;
         }
         state = state_;
         return buf;
       }
+      uint8_t prev[kStride + 1];
+      prev[0] = state_;
       for (int i = 0; i < kStride; ++i) {
         uint64_t row = num_dfa_table[uint8_t(*buf)];
-        auto next = row >> (state_ & 63);
-        if ((next & 63) == 0) {
-          state = state_;
+        prev[i + 1] = row >> (prev[i] & 63);
+        if ((prev[i + 1] & 63) == 0) {
+          state = prev[i];
           return buf;
         }
-        state_ = next;
         ++buf;
       }
+      state_ = prev[kStride];
     }
   }