Improve regex vm to next start performance by avoiding iterator copies

src/regex_impl.hh

@@ -525,10 +525,15 @@ private:
 
     void to_next_start(EffectiveIt& start, const EffectiveIt& end, const StartDesc& start_desc)
     {
-        Codepoint cp;
+        while (start != end)
-        while (start != end and (cp = *start) >= 0 and
+        {
-               not start_desc.map[cp < StartDesc::count ? cp : StartDesc::other])
+            const Codepoint cp = read(start);
-            ++start;
+            if (start_desc.map[(cp >= 0 and cp < StartDesc::count) ? cp : StartDesc::other])
+            {
+                --start;
+                return;
+            }
+        }
     }
 
     template<MatchDirection look_direction, bool ignore_case>
@@ -596,6 +601,9 @@ private:
         return is_word(*(pos-1)) != is_word(*pos);
     }
 
+    static Codepoint read(Utf8It& it) { return it.read(); }
+    static Codepoint read(std::reverse_iterator<Utf8It>& it) { Codepoint cp = *it; ++it; return cp; }
+
     static const Iterator& get_base(const Utf8It& it) { return it.base(); }
     static Iterator get_base(const std::reverse_iterator<Utf8It>& it) { return it.base().base(); }
 

src/utf8_iterator.hh

@@ -130,6 +130,11 @@ public:
         return get_value();
     }
 
+    CodepointType read() noexcept(noexcept_policy)
+    {
+        return (CodepointType)utf8::read_codepoint<InvalidPolicy>(m_it, m_end);
+    }
+
     const BaseIt& base() const noexcept(noexcept_policy) { return m_it; }
 
 private: