Changeset 1055 for XmlTools2/trunk/libs/pugixml.cpp

Timestamp:

Oct 28, 2016, 11:05:06 PM (8 years ago)

Author:

s10k

Message:

XmlTools 2.0c

File:

• XmlTools2/trunk/libs/pugixml.cpp (modified) (328 diffs)

XmlTools2/trunk/libs/pugixml.cpp

@@ -1 +1 @@
 /**
- * pugixml parser - version 1.2
+ * pugixml parser - version 1.7
  * --------------------------------------------------------
- * Copyright (C) 2006-2012, by Arseny Kapoulkine (arseny.kapoulkine@gmail.com)
+ * Copyright (C) 2006-2015, by Arseny Kapoulkine (arseny.kapoulkine@gmail.com)
  * Report bugs and download new versions at http://pugixml.org/
  *
@@ -21 +21 @@
 #include <string.h>
 #include <assert.h>
-#include <wchar.h>
+#include <limits.h>
+
+#ifdef PUGIXML_WCHAR_MODE
+#       include <wchar.h>
+#endif
 
 #ifndef PUGIXML_NO_XPATH
@@ -82 +86 @@
 #endif
 
+// Branch weight controls
+#if defined(__GNUC__)
+#       define PUGI__UNLIKELY(cond) __builtin_expect(cond, 0)
+#else
+#       define PUGI__UNLIKELY(cond) (cond)
+#endif
+
 // Simple static assertion
 #define PUGI__STATIC_ASSERT(cond) { static const char condition_failed[(cond) ? 1 : -1] = {0}; (void)condition_failed[0]; }
@@ -96 +107 @@
 using std::memcpy;
 using std::memmove;
+using std::memset;
 #endif
 
@@ -124 +136 @@
 #       include <stdint.h>
 #else
+namespace pugi
+{
 #       ifndef _UINTPTR_T_DEFINED
-// No native uintptr_t in MSVC6 and in some WinCE versions
-typedef size_t uintptr_t;
-#define _UINTPTR_T_DEFINED
+        typedef size_t uintptr_t;
 #       endif
-PUGI__NS_BEGIN
+
         typedef unsigned __int8 uint8_t;
         typedef unsigned __int16 uint16_t;
         typedef unsigned __int32 uint32_t;
-PUGI__NS_END
+}
 #endif
 
@@ -155 +167 @@
         };
 
+        // Global allocation functions are stored in class statics so that in header mode linker deduplicates them
+        // Without a template<> we'll get multiple definitions of the same static
         template <typename T> allocation_function xml_memory_management_function_storage<T>::allocate = default_allocate;
         template <typename T> deallocation_function xml_memory_management_function_storage<T>::deallocate = default_deallocate;
@@ -196 +210 @@
                 return lhs[count] == 0;
         }
-        
-#ifdef PUGIXML_WCHAR_MODE
-        // Convert string to wide string, assuming all symbols are ASCII
-        PUGI__FN void widen_ascii(wchar_t* dest, const char* source)
-        {
-                for (const char* i = source; *i; ++i) *dest++ = *i;
-                *dest = 0;
-        }
-#endif
+
+        // Get length of wide string, even if CRT lacks wide character support
+        PUGI__FN size_t strlength_wide(const wchar_t* s)
+        {
+                assert(s);
+
+        #ifdef PUGIXML_WCHAR_MODE
+                return wcslen(s);
+        #else
+                const wchar_t* end = s;
+                while (*end) end++;
+                return static_cast<size_t>(end - s);
+        #endif
+        }
 PUGI__NS_END
 
-#if !defined(PUGIXML_NO_STL) || !defined(PUGIXML_NO_XPATH)
-// auto_ptr-like buffer holder for exception recovery
+// auto_ptr-like object for exception recovery
 PUGI__NS_BEGIN
-        struct buffer_holder
-        {
-                void* data;
-                void (*deleter)(void*);
-
-                buffer_holder(void* data_, void (*deleter_)(void*)): data(data_), deleter(deleter_)
-                {
-                }
-
-                ~buffer_holder()
+        template <typename T, typename D = void(*)(T*)> struct auto_deleter
+        {
+                T* data;
+                D deleter;
+
+                auto_deleter(T* data_, D deleter_): data(data_), deleter(deleter_)
+                {
+                }
+
+                ~auto_deleter()
                 {
                         if (data) deleter(data);
                 }
 
-                void* release()
-                {
-                        void* result = data;
+                T* release()
+                {
+                        T* result = data;
                         data = 0;
                         return result;
                 }
         };
+PUGI__NS_END
+
+#ifdef PUGIXML_COMPACT
+PUGI__NS_BEGIN
+        class compact_hash_table
+        {
+        public:
+                compact_hash_table(): _items(0), _capacity(0), _count(0)
+                {
+                }
+
+                void clear()
+                {
+                        if (_items)
+                        {
+                                xml_memory::deallocate(_items);
+                                _items = 0;
+                                _capacity = 0;
+                                _count = 0;
+                        }
+                }
+
+                void** find(const void* key)
+                {
+                        assert(key);
+
+                        if (_capacity == 0) return 0;
+
+                        size_t hashmod = _capacity - 1;
+                        size_t bucket = hash(key) & hashmod;
+
+                        for (size_t probe = 0; probe <= hashmod; ++probe)
+                        {
+                                item_t& probe_item = _items[bucket];
+
+                                if (probe_item.key == key)
+                                        return &probe_item.value;
+
+                                if (probe_item.key == 0)
+                                        return 0;
+
+                                // hash collision, quadratic probing
+                                bucket = (bucket + probe + 1) & hashmod;
+                        }
+
+                        assert(!"Hash table is full");
+                        return 0;
+                }
+
+                void** insert(const void* key)
+                {
+                        assert(key);
+                        assert(_capacity != 0 && _count < _capacity - _capacity / 4);
+
+                        size_t hashmod = _capacity - 1;
+                        size_t bucket = hash(key) & hashmod;
+
+                        for (size_t probe = 0; probe <= hashmod; ++probe)
+                        {
+                                item_t& probe_item = _items[bucket];
+
+                                if (probe_item.key == 0)
+                                {
+                                        probe_item.key = key;
+                                        _count++;
+                                        return &probe_item.value;
+                                }
+
+                                if (probe_item.key == key)
+                                        return &probe_item.value;
+
+                                // hash collision, quadratic probing
+                                bucket = (bucket + probe + 1) & hashmod;
+                        }
+
+                        assert(!"Hash table is full");
+                        return 0;
+                }
+
+                bool reserve()
+                {
+                        if (_count + 16 >= _capacity - _capacity / 4)
+                                return rehash();
+
+                        return true;
+                }
+
+        private:
+                struct item_t
+                {
+                        const void* key;
+                        void* value;
+                };
+
+                item_t* _items;
+                size_t _capacity;
+
+                size_t _count;
+
+                bool rehash();
+
+                static unsigned int hash(const void* key)
+                {
+                        unsigned int h = static_cast<unsigned int>(reinterpret_cast<uintptr_t>(key));
+
+                        // MurmurHash3 32-bit finalizer
+                        h ^= h >> 16;
+                        h *= 0x85ebca6bu;
+                        h ^= h >> 13;
+                        h *= 0xc2b2ae35u;
+                        h ^= h >> 16;
+
+                        return h;
+                }
+        };
+
+        PUGI__FN_NO_INLINE bool compact_hash_table::rehash()
+        {
+                compact_hash_table rt;
+                rt._capacity = (_capacity == 0) ? 32 : _capacity * 2;
+                rt._items = static_cast<item_t*>(xml_memory::allocate(sizeof(item_t) * rt._capacity));
+
+                if (!rt._items)
+                        return false;
+
+                memset(rt._items, 0, sizeof(item_t) * rt._capacity);
+
+                for (size_t i = 0; i < _capacity; ++i)
+                        if (_items[i].key)
+                                *rt.insert(_items[i].key) = _items[i].value;
+
+                if (_items)
+                        xml_memory::deallocate(_items);
+
+                _capacity = rt._capacity;
+                _items = rt._items;
+
+                assert(_count == rt._count);
+
+                return true;
+        }
+
 PUGI__NS_END
 #endif
@@ -243 +403 @@
                 ;
 
-        static const uintptr_t xml_memory_page_alignment = 32;
+#ifdef PUGIXML_COMPACT
+        static const uintptr_t xml_memory_block_alignment = 4;
+
+        static const uintptr_t xml_memory_page_alignment = sizeof(void*);
+#else
+        static const uintptr_t xml_memory_block_alignment = sizeof(void*);
+
+        static const uintptr_t xml_memory_page_alignment = 64;
         static const uintptr_t xml_memory_page_pointer_mask = ~(xml_memory_page_alignment - 1);
+#endif
+
+        // extra metadata bits
+        static const uintptr_t xml_memory_page_contents_shared_mask = 32;
         static const uintptr_t xml_memory_page_name_allocated_mask = 16;
         static const uintptr_t xml_memory_page_value_allocated_mask = 8;
         static const uintptr_t xml_memory_page_type_mask = 7;
 
+        // combined masks for string uniqueness
+        static const uintptr_t xml_memory_page_name_allocated_or_shared_mask = xml_memory_page_name_allocated_mask | xml_memory_page_contents_shared_mask;
+        static const uintptr_t xml_memory_page_value_allocated_or_shared_mask = xml_memory_page_value_allocated_mask | xml_memory_page_contents_shared_mask;
+
+#ifdef PUGIXML_COMPACT
+        #define PUGI__GETPAGE_IMPL(header) (header).get_page()
+#else
+        #define PUGI__GETPAGE_IMPL(header) reinterpret_cast<impl::xml_memory_page*>((header) & impl::xml_memory_page_pointer_mask)
+#endif
+
+        #define PUGI__GETPAGE(n) PUGI__GETPAGE_IMPL((n)->header)
+        #define PUGI__NODETYPE(n) static_cast<xml_node_type>(((n)->header & impl::xml_memory_page_type_mask) + 1)
+
         struct xml_allocator;
 
@@ -255 +439 @@
                 static xml_memory_page* construct(void* memory)
                 {
-                        if (!memory) return 0; //$ redundant, left for performance
-
                         xml_memory_page* result = static_cast<xml_memory_page*>(memory);
 
                         result->allocator = 0;
-                        result->memory = 0;
                         result->prev = 0;
                         result->next = 0;
@@ -266 +447 @@
                         result->freed_size = 0;
 
+                #ifdef PUGIXML_COMPACT
+                        result->compact_string_base = 0;
+                        result->compact_shared_parent = 0;
+                        result->compact_page_marker = 0;
+                #endif
+
                         return result;
                 }
 
                 xml_allocator* allocator;
-
-                void* memory;
 
                 xml_memory_page* prev;
@@ -279 +464 @@
                 size_t freed_size;
 
-                char data[1];
+        #ifdef PUGIXML_COMPACT
+                char_t* compact_string_base;
+                void* compact_shared_parent;
+                uint32_t* compact_page_marker;
+        #endif
         };
 
@@ -292 +481 @@
                 xml_allocator(xml_memory_page* root): _root(root), _busy_size(root->busy_size)
                 {
+                #ifdef PUGIXML_COMPACT
+                        _hash = 0;
+                #endif
                 }
 
                 xml_memory_page* allocate_page(size_t data_size)
                 {
-                        size_t size = offsetof(xml_memory_page, data) + data_size;
+                        size_t size = sizeof(xml_memory_page) + data_size;
 
                         // allocate block with some alignment, leaving memory for worst-case padding
@@ -302 +494 @@
                         if (!memory) return 0;
 
-                        // align upwards to page boundary
-                        void* page_memory = reinterpret_cast<void*>((reinterpret_cast<uintptr_t>(memory) + (xml_memory_page_alignment - 1)) & ~(xml_memory_page_alignment - 1));
+                        // align to next page boundary (note: this guarantees at least 1 usable byte before the page)
+                        char* page_memory = reinterpret_cast<char*>((reinterpret_cast<uintptr_t>(memory) + xml_memory_page_alignment) & ~(xml_memory_page_alignment - 1));
 
                         // prepare page structure
                         xml_memory_page* page = xml_memory_page::construct(page_memory);
-
-                        page->memory = memory;
+                        assert(page);
+
                         page->allocator = _root->allocator;
 
+                        // record the offset for freeing the memory block
+                        assert(page_memory > memory && page_memory - static_cast<char*>(memory) <= 127);
+                        page_memory[-1] = static_cast<char>(page_memory - static_cast<char*>(memory));
+
                         return page;
                 }
@@ -316 +512 @@
                 static void deallocate_page(xml_memory_page* page)
                 {
-                        xml_memory::deallocate(page->memory);
+                        char* page_memory = reinterpret_cast<char*>(page);
+
+                        xml_memory::deallocate(page_memory - page_memory[-1]);
                 }
 
@@ -323 +521 @@
                 void* allocate_memory(size_t size, xml_memory_page*& out_page)
                 {
-                        if (_busy_size + size > xml_memory_page_size) return allocate_memory_oob(size, out_page);
-
-                        void* buf = _root->data + _busy_size;
+                        if (PUGI__UNLIKELY(_busy_size + size > xml_memory_page_size))
+                                return allocate_memory_oob(size, out_page);
+
+                        void* buf = reinterpret_cast<char*>(_root) + sizeof(xml_memory_page) + _busy_size;
 
                         _busy_size += size;
@@ -334 +533 @@
                 }
 
+        #ifdef PUGIXML_COMPACT
+                void* allocate_object(size_t size, xml_memory_page*& out_page)
+                {
+                        void* result = allocate_memory(size + sizeof(uint32_t), out_page);
+                        if (!result) return 0;
+
+                        // adjust for marker
+                        ptrdiff_t offset = static_cast<char*>(result) - reinterpret_cast<char*>(out_page->compact_page_marker);
+
+                        if (PUGI__UNLIKELY(static_cast<uintptr_t>(offset) >= 256 * xml_memory_block_alignment))
+                        {
+                                // insert new marker
+                                uint32_t* marker = static_cast<uint32_t*>(result);
+
+                                *marker = static_cast<uint32_t>(reinterpret_cast<char*>(marker) - reinterpret_cast<char*>(out_page));
+                                out_page->compact_page_marker = marker;
+
+                                // since we don't reuse the page space until we reallocate it, we can just pretend that we freed the marker block
+                                // this will make sure deallocate_memory correctly tracks the size
+                                out_page->freed_size += sizeof(uint32_t);
+
+                                return marker + 1;
+                        }
+                        else
+                        {
+                                // roll back uint32_t part
+                                _busy_size -= sizeof(uint32_t);
+
+                                return result;
+                        }
+                }
+        #else
+                void* allocate_object(size_t size, xml_memory_page*& out_page)
+                {
+                        return allocate_memory(size, out_page);
+                }
+        #endif
+
                 void deallocate_memory(void* ptr, size_t size, xml_memory_page* page)
                 {
                         if (page == _root) page->busy_size = _busy_size;
 
-                        assert(ptr >= page->data && ptr < page->data + page->busy_size);
+                        assert(ptr >= reinterpret_cast<char*>(page) + sizeof(xml_memory_page) && ptr < reinterpret_cast<char*>(page) + sizeof(xml_memory_page) + page->busy_size);
                         (void)!ptr;
 
@@ -351 +588 @@
 
                                         // top page freed, just reset sizes
-                                        page->busy_size = page->freed_size = 0;
+                                        page->busy_size = 0;
+                                        page->freed_size = 0;
+
+                                #ifdef PUGIXML_COMPACT
+                                        // reset compact state to maximize efficiency
+                                        page->compact_string_base = 0;
+                                        page->compact_shared_parent = 0;
+                                        page->compact_page_marker = 0;
+                                #endif
+
                                         _busy_size = 0;
                                 }
@@ -371 +617 @@
                 char_t* allocate_string(size_t length)
                 {
+                        static const size_t max_encoded_offset = (1 << 16) * xml_memory_block_alignment;
+
+                        PUGI__STATIC_ASSERT(xml_memory_page_size <= max_encoded_offset);
+
                         // allocate memory for string and header block
                         size_t size = sizeof(xml_memory_string_header) + length * sizeof(char_t);
                         
-                        // round size up to pointer alignment boundary
-                        size_t full_size = (size + (sizeof(void*) - 1)) & ~(sizeof(void*) - 1);
+                        // round size up to block alignment boundary
+                        size_t full_size = (size + (xml_memory_block_alignment - 1)) & ~(xml_memory_block_alignment - 1);
 
                         xml_memory_page* page;
@@ -383 +633 @@
 
                         // setup header
-                        ptrdiff_t page_offset = reinterpret_cast<char*>(header) - page->data;
-
-                        assert(page_offset >= 0 && page_offset < (1 << 16));
-                        header->page_offset = static_cast<uint16_t>(page_offset);
+                        ptrdiff_t page_offset = reinterpret_cast<char*>(header) - reinterpret_cast<char*>(page) - sizeof(xml_memory_page);
+
+                        assert(page_offset % xml_memory_block_alignment == 0);
+                        assert(page_offset >= 0 && static_cast<size_t>(page_offset) < max_encoded_offset);
+                        header->page_offset = static_cast<uint16_t>(static_cast<size_t>(page_offset) / xml_memory_block_alignment);
 
                         // full_size == 0 for large strings that occupy the whole page
-                        assert(full_size < (1 << 16) || (page->busy_size == full_size && page_offset == 0));
-                        header->full_size = static_cast<uint16_t>(full_size < (1 << 16) ? full_size : 0);
+                        assert(full_size % xml_memory_block_alignment == 0);
+                        assert(full_size < max_encoded_offset || (page->busy_size == full_size && page_offset == 0));
+                        header->full_size = static_cast<uint16_t>(full_size < max_encoded_offset ? full_size / xml_memory_block_alignment : 0);
 
                         // round-trip through void* to avoid 'cast increases required alignment of target type' warning
@@ -404 +656 @@
                         // get header
                         xml_memory_string_header* header = static_cast<xml_memory_string_header*>(static_cast<void*>(string)) - 1;
+                        assert(header);
 
                         // deallocate
-                        size_t page_offset = offsetof(xml_memory_page, data) + header->page_offset;
+                        size_t page_offset = sizeof(xml_memory_page) + header->page_offset * xml_memory_block_alignment;
                         xml_memory_page* page = reinterpret_cast<xml_memory_page*>(static_cast<void*>(reinterpret_cast<char*>(header) - page_offset));
 
                         // if full_size == 0 then this string occupies the whole page
-                        size_t full_size = header->full_size == 0 ? page->busy_size : header->full_size;
+                        size_t full_size = header->full_size == 0 ? page->busy_size : header->full_size * xml_memory_block_alignment;
 
                         deallocate_memory(header, full_size, page);
+                }
+
+                bool reserve()
+                {
+                #ifdef PUGIXML_COMPACT
+                        return _hash->reserve();
+                #else
+                        return true;
+                #endif
                 }
 
                 xml_memory_page* _root;
                 size_t _busy_size;
+
+        #ifdef PUGIXML_COMPACT
+                compact_hash_table* _hash;
+        #endif
         };
 
@@ -450 +716 @@
                         _root->prev->next = page;
                         _root->prev = page;
-                }
-
-                // allocate inside page
-                page->busy_size = size;
-
-                return page->data;
+
+                        page->busy_size = size;
+                }
+
+                return reinterpret_cast<char*>(page) + sizeof(xml_memory_page);
         }
 PUGI__NS_END
 
+#ifdef PUGIXML_COMPACT
+PUGI__NS_BEGIN
+        static const uintptr_t compact_alignment_log2 = 2;
+        static const uintptr_t compact_alignment = 1 << compact_alignment_log2;
+
+        class compact_header
+        {
+        public:
+                compact_header(xml_memory_page* page, unsigned int flags)
+                {
+                        PUGI__STATIC_ASSERT(xml_memory_block_alignment == compact_alignment);
+
+                        ptrdiff_t offset = (reinterpret_cast<char*>(this) - reinterpret_cast<char*>(page->compact_page_marker));
+                        assert(offset % compact_alignment == 0 && static_cast<uintptr_t>(offset) < 256 * compact_alignment);
+
+                        _page = static_cast<unsigned char>(offset >> compact_alignment_log2);
+                        _flags = static_cast<unsigned char>(flags);
+                }
+
+                void operator&=(uintptr_t mod)
+                {
+                        _flags &= static_cast<unsigned char>(mod);
+                }
+
+                void operator|=(uintptr_t mod)
+                {
+                        _flags |= static_cast<unsigned char>(mod);
+                }
+
+                uintptr_t operator&(uintptr_t mod) const
+                {
+                        return _flags & mod;
+                }
+
+                xml_memory_page* get_page() const
+                {
+                        // round-trip through void* to silence 'cast increases required alignment of target type' warnings
+                        const char* page_marker = reinterpret_cast<const char*>(this) - (_page << compact_alignment_log2);
+                        const char* page = page_marker - *reinterpret_cast<const uint32_t*>(static_cast<const void*>(page_marker));
+
+                        return const_cast<xml_memory_page*>(reinterpret_cast<const xml_memory_page*>(static_cast<const void*>(page)));
+                }
+
+        private:
+                unsigned char _page;
+                unsigned char _flags;
+        };
+
+        PUGI__FN xml_memory_page* compact_get_page(const void* object, int header_offset)
+        {
+                const compact_header* header = reinterpret_cast<const compact_header*>(static_cast<const char*>(object) - header_offset);
+
+                return header->get_page();
+        }
+
+        template <int header_offset, typename T> PUGI__FN_NO_INLINE T* compact_get_value(const void* object)
+        {
+                return static_cast<T*>(*compact_get_page(object, header_offset)->allocator->_hash->find(object));
+        }
+
+        template <int header_offset, typename T> PUGI__FN_NO_INLINE void compact_set_value(const void* object, T* value)
+        {
+                *compact_get_page(object, header_offset)->allocator->_hash->insert(object) = value;
+        }
+
+        template <typename T, int header_offset, int start = -126> class compact_pointer
+        {
+        public:
+                compact_pointer(): _data(0)
+                {
+                }
+
+                void operator=(const compact_pointer& rhs)
+                {
+                        *this = rhs + 0;
+                }
+
+                void operator=(T* value)
+                {
+                        if (value)
+                        {
+                                // value is guaranteed to be compact-aligned; 'this' is not
+                                // our decoding is based on 'this' aligned to compact alignment downwards (see operator T*)
+                                // so for negative offsets (e.g. -3) we need to adjust the diff by compact_alignment - 1 to
+                                // compensate for arithmetic shift rounding for negative values
+                                ptrdiff_t diff = reinterpret_cast<char*>(value) - reinterpret_cast<char*>(this);
+                                ptrdiff_t offset = ((diff + int(compact_alignment - 1)) >> compact_alignment_log2) - start;
+
+                                if (static_cast<uintptr_t>(offset) <= 253)
+                                        _data = static_cast<unsigned char>(offset + 1);
+                                else
+                                {
+                                        compact_set_value<header_offset>(this, value);
+
+                                        _data = 255;
+                                }
+                        }
+                        else
+                                _data = 0;
+                }
+
+                operator T*() const
+                {
+                        if (_data)
+                        {
+                                if (_data < 255)
+                                {
+                                        uintptr_t base = reinterpret_cast<uintptr_t>(this) & ~(compact_alignment - 1);
+
+                                        return reinterpret_cast<T*>(base + ((_data - 1 + start) << compact_alignment_log2));
+                                }
+                                else
+                                        return compact_get_value<header_offset, T>(this);
+                        }
+                        else
+                                return 0;
+                }
+
+                T* operator->() const
+                {
+                        return *this;
+                }
+
+        private:
+                unsigned char _data;
+        };
+
+        template <typename T, int header_offset> class compact_pointer_parent
+        {
+        public:
+                compact_pointer_parent(): _data(0)
+                {
+                }
+
+                void operator=(const compact_pointer_parent& rhs)
+                {
+                        *this = rhs + 0;
+                }
+
+                void operator=(T* value)
+                {
+                        if (value)
+                        {
+                                // value is guaranteed to be compact-aligned; 'this' is not
+                                // our decoding is based on 'this' aligned to compact alignment downwards (see operator T*)
+                                // so for negative offsets (e.g. -3) we need to adjust the diff by compact_alignment - 1 to
+                                // compensate for arithmetic shift behavior for negative values
+                                ptrdiff_t diff = reinterpret_cast<char*>(value) - reinterpret_cast<char*>(this);
+                                ptrdiff_t offset = ((diff + int(compact_alignment - 1)) >> compact_alignment_log2) + 65533;
+
+                                if (static_cast<uintptr_t>(offset) <= 65533)
+                                {
+                                        _data = static_cast<unsigned short>(offset + 1);
+                                }
+                                else
+                                {
+                                        xml_memory_page* page = compact_get_page(this, header_offset);
+
+                                        if (PUGI__UNLIKELY(page->compact_shared_parent == 0))
+                                                page->compact_shared_parent = value;
+
+                                        if (page->compact_shared_parent == value)
+                                        {
+                                                _data = 65534;
+                                        }
+                                        else
+                                        {
+                                                compact_set_value<header_offset>(this, value);
+
+                                                _data = 65535;
+                                        }
+                                }
+                        }
+                        else
+                        {
+                                _data = 0;
+                        }
+                }
+
+                operator T*() const
+                {
+                        if (_data)
+                        {
+                                if (_data < 65534)
+                                {
+                                        uintptr_t base = reinterpret_cast<uintptr_t>(this) & ~(compact_alignment - 1);
+
+                                        return reinterpret_cast<T*>(base + ((_data - 1 - 65533) << compact_alignment_log2));
+                                }
+                                else if (_data == 65534)
+                                        return static_cast<T*>(compact_get_page(this, header_offset)->compact_shared_parent);
+                                else
+                                        return compact_get_value<header_offset, T>(this);
+                        }
+                        else
+                                return 0;
+                }
+
+                T* operator->() const
+                {
+                        return *this;
+                }
+
+        private:
+                uint16_t _data;
+        };
+
+        template <int header_offset, int base_offset> class compact_string
+        {
+        public:
+                compact_string(): _data(0)
+                {
+                }
+
+                void operator=(const compact_string& rhs)
+                {
+                        *this = rhs + 0;
+                }
+
+                void operator=(char_t* value)
+                {
+                        if (value)
+                        {
+                                xml_memory_page* page = compact_get_page(this, header_offset);
+
+                                if (PUGI__UNLIKELY(page->compact_string_base == 0))
+                                        page->compact_string_base = value;
+
+                                ptrdiff_t offset = value - page->compact_string_base;
+
+                                if (static_cast<uintptr_t>(offset) < (65535 << 7))
+                                {
+                                        // round-trip through void* to silence 'cast increases required alignment of target type' warnings
+                                        uint16_t* base = reinterpret_cast<uint16_t*>(static_cast<void*>(reinterpret_cast<char*>(this) - base_offset));
+
+                                        if (*base == 0)
+                                        {
+                                                *base = static_cast<uint16_t>((offset >> 7) + 1);
+                                                _data = static_cast<unsigned char>((offset & 127) + 1);
+                                        }
+                                        else
+                                        {
+                                                ptrdiff_t remainder = offset - ((*base - 1) << 7);
+
+                                                if (static_cast<uintptr_t>(remainder) <= 253)
+                                                {
+                                                        _data = static_cast<unsigned char>(remainder + 1);
+                                                }
+                                                else
+                                                {
+                                                        compact_set_value<header_offset>(this, value);
+
+                                                        _data = 255;
+                                                }
+                                        }
+                                }
+                                else
+                                {
+                                        compact_set_value<header_offset>(this, value);
+
+                                        _data = 255;
+                                }
+                        }
+                        else
+                        {
+                                _data = 0;
+                        }
+                }
+
+                operator char_t*() const
+                {
+                        if (_data)
+                        {
+                                if (_data < 255)
+                                {
+                                        xml_memory_page* page = compact_get_page(this, header_offset);
+
+                                        // round-trip through void* to silence 'cast increases required alignment of target type' warnings
+                                        const uint16_t* base = reinterpret_cast<const uint16_t*>(static_cast<const void*>(reinterpret_cast<const char*>(this) - base_offset));
+                                        assert(*base);
+
+                                        ptrdiff_t offset = ((*base - 1) << 7) + (_data - 1);
+
+                                        return page->compact_string_base + offset;
+                                }
+                                else
+                                {
+                                        return compact_get_value<header_offset, char_t>(this);
+                                }
+                        }
+                        else
+                                return 0;
+                }
+
+        private:
+                unsigned char _data;
+        };
+PUGI__NS_END
+#endif
+
+#ifdef PUGIXML_COMPACT
 namespace pugi
 {
-        /// A 'name=value' XML attribute structure.
         struct xml_attribute_struct
         {
-                /// Default ctor
-                xml_attribute_struct(impl::xml_memory_page* page): header(reinterpret_cast<uintptr_t>(page)), name(0), value(0), prev_attribute_c(0), next_attribute(0)
-                {
-                }
-
-                uintptr_t header;
-
-                char_t* name;   ///< Pointer to attribute name.
-                char_t* value;  ///< Pointer to attribute value.
-
-                xml_attribute_struct* prev_attribute_c; ///< Previous attribute (cyclic list)
-                xml_attribute_struct* next_attribute;   ///< Next attribute
+                xml_attribute_struct(impl::xml_memory_page* page): header(page, 0), namevalue_base(0)
+                {
+                        PUGI__STATIC_ASSERT(sizeof(xml_attribute_struct) == 8);
+                }
+
+                impl::compact_header header;
+
+                uint16_t namevalue_base;
+
+                impl::compact_string<4, 2> name;
+                impl::compact_string<5, 3> value;
+
+                impl::compact_pointer<xml_attribute_struct, 6> prev_attribute_c;
+                impl::compact_pointer<xml_attribute_struct, 7, 0> next_attribute;
         };
 
-        /// An XML document tree node.
         struct xml_node_struct
         {
-                /// Default ctor
-                /// \param type - node type
-                xml_node_struct(impl::xml_memory_page* page, xml_node_type type): header(reinterpret_cast<uintptr_t>(page) | (type - 1)), parent(0), name(0), value(0), first_child(0), prev_sibling_c(0), next_sibling(0), first_attribute(0)
-                {
-                }
-
-                uintptr_t header;
-
-                xml_node_struct*                parent;                                 ///< Pointer to parent
-
-                char_t*                                 name;                                   ///< Pointer to element name.
-                char_t*                                 value;                                  ///< Pointer to any associated string data.
-
-                xml_node_struct*                first_child;                    ///< First child
-                
-                xml_node_struct*                prev_sibling_c;                 ///< Left brother (cyclic list)
-                xml_node_struct*                next_sibling;                   ///< Right brother
-                
-                xml_attribute_struct*   first_attribute;                ///< First attribute
+                xml_node_struct(impl::xml_memory_page* page, xml_node_type type): header(page, type - 1), namevalue_base(0)
+                {
+                        PUGI__STATIC_ASSERT(sizeof(xml_node_struct) == 12);
+                }
+
+                impl::compact_header header;
+
+                uint16_t namevalue_base;
+
+                impl::compact_string<4, 2> name;
+                impl::compact_string<5, 3> value;
+
+                impl::compact_pointer_parent<xml_node_struct, 6> parent;
+
+                impl::compact_pointer<xml_node_struct, 8, 0> first_child;
+
+                impl::compact_pointer<xml_node_struct,  9>    prev_sibling_c;
+                impl::compact_pointer<xml_node_struct, 10, 0> next_sibling;
+
+                impl::compact_pointer<xml_attribute_struct, 11, 0> first_attribute;
         };
 }
+#else
+namespace pugi
+{
+        struct xml_attribute_struct
+        {
+                xml_attribute_struct(impl::xml_memory_page* page): header(reinterpret_cast<uintptr_t>(page)), name(0), value(0), prev_attribute_c(0), next_attribute(0)
+                {
+                }
+
+                uintptr_t header;
+
+                char_t* name;
+                char_t* value;
+
+                xml_attribute_struct* prev_attribute_c;
+                xml_attribute_struct* next_attribute;
+        };
+
+        struct xml_node_struct
+        {
+                xml_node_struct(impl::xml_memory_page* page, xml_node_type type): header(reinterpret_cast<uintptr_t>(page) | (type - 1)), name(0), value(0), parent(0), first_child(0), prev_sibling_c(0), next_sibling(0), first_attribute(0)
+                {
+                }
+
+                uintptr_t header;
+
+                char_t* name;
+                char_t* value;
+
+                xml_node_struct* parent;
+
+                xml_node_struct* first_child;
+
+                xml_node_struct* prev_sibling_c;
+                xml_node_struct* next_sibling;
+
+                xml_attribute_struct* first_attribute;
+        };
+}
+#endif
 
 PUGI__NS_BEGIN
+        struct xml_extra_buffer
+        {
+                char_t* buffer;
+                xml_extra_buffer* next;
+        };
+
         struct xml_document_struct: public xml_node_struct, public xml_allocator
         {
-                xml_document_struct(xml_memory_page* page): xml_node_struct(page, node_document), xml_allocator(page), buffer(0)
-                {
+                xml_document_struct(xml_memory_page* page): xml_node_struct(page, node_document), xml_allocator(page), buffer(0), extra_buffers(0)
+                {
+                #ifdef PUGIXML_COMPACT
+                        _hash = &hash;
+                #endif
                 }
 
                 const char_t* buffer;
+
+                xml_extra_buffer* extra_buffers;
+
+        #ifdef PUGIXML_COMPACT
+                compact_hash_table hash;
+        #endif
         };
 
-        inline xml_allocator& get_allocator(const xml_node_struct* node)
-        {
-                assert(node);
-
-                return *reinterpret_cast<xml_memory_page*>(node->header & xml_memory_page_pointer_mask)->allocator;
+        template <typename Object> inline xml_allocator& get_allocator(const Object* object)
+        {
+                assert(object);
+
+                return *PUGI__GETPAGE(object)->allocator;
+        }
+
+        template <typename Object> inline xml_document_struct& get_document(const Object* object)
+        {
+                assert(object);
+
+                return *static_cast<xml_document_struct*>(PUGI__GETPAGE(object)->allocator);
         }
 PUGI__NS_END
@@ -526 +1155 @@
         {
                 xml_memory_page* page;
-                void* memory = alloc.allocate_memory(sizeof(xml_attribute_struct), page);
+                void* memory = alloc.allocate_object(sizeof(xml_attribute_struct), page);
+                if (!memory) return 0;
 
                 return new (memory) xml_attribute_struct(page);
@@ -534 +1164 @@
         {
                 xml_memory_page* page;
-                void* memory = alloc.allocate_memory(sizeof(xml_node_struct), page);
+                void* memory = alloc.allocate_object(sizeof(xml_node_struct), page);
+                if (!memory) return 0;
 
                 return new (memory) xml_node_struct(page, type);
@@ -541 +1172 @@
         inline void destroy_attribute(xml_attribute_struct* a, xml_allocator& alloc)
         {
-                uintptr_t header = a->header;
-
-                if (header & impl::xml_memory_page_name_allocated_mask) alloc.deallocate_string(a->name);
-                if (header & impl::xml_memory_page_value_allocated_mask) alloc.deallocate_string(a->value);
-
-                alloc.deallocate_memory(a, sizeof(xml_attribute_struct), reinterpret_cast<xml_memory_page*>(header & xml_memory_page_pointer_mask));
+                if (a->header & impl::xml_memory_page_name_allocated_mask)
+                        alloc.deallocate_string(a->name);
+
+                if (a->header & impl::xml_memory_page_value_allocated_mask)
+                        alloc.deallocate_string(a->value);
+
+                alloc.deallocate_memory(a, sizeof(xml_attribute_struct), PUGI__GETPAGE(a));
         }
 
         inline void destroy_node(xml_node_struct* n, xml_allocator& alloc)
         {
-                uintptr_t header = n->header;
-
-                if (header & impl::xml_memory_page_name_allocated_mask) alloc.deallocate_string(n->name);
-                if (header & impl::xml_memory_page_value_allocated_mask) alloc.deallocate_string(n->value);
+                if (n->header & impl::xml_memory_page_name_allocated_mask)
+                        alloc.deallocate_string(n->name);
+
+                if (n->header & impl::xml_memory_page_value_allocated_mask)
+                        alloc.deallocate_string(n->value);
 
                 for (xml_attribute_struct* attr = n->first_attribute; attr; )
@@ -574 +1207 @@
                 }
 
-                alloc.deallocate_memory(n, sizeof(xml_node_struct), reinterpret_cast<xml_memory_page*>(header & xml_memory_page_pointer_mask));
-        }
-
-        PUGI__FN_NO_INLINE xml_node_struct* append_node(xml_node_struct* node, xml_allocator& alloc, xml_node_type type = node_element)
-        {
+                alloc.deallocate_memory(n, sizeof(xml_node_struct), PUGI__GETPAGE(n));
+        }
+
+        inline void append_node(xml_node_struct* child, xml_node_struct* node)
+        {
+                child->parent = node;
+
+                xml_node_struct* head = node->first_child;
+
+                if (head)
+                {
+                        xml_node_struct* tail = head->prev_sibling_c;
+
+                        tail->next_sibling = child;
+                        child->prev_sibling_c = tail;
+                        head->prev_sibling_c = child;
+                }
+                else
+                {
+                        node->first_child = child;
+                        child->prev_sibling_c = child;
+                }
+        }
+
+        inline void prepend_node(xml_node_struct* child, xml_node_struct* node)
+        {
+                child->parent = node;
+
+                xml_node_struct* head = node->first_child;
+
+                if (head)
+                {
+                        child->prev_sibling_c = head->prev_sibling_c;
+                        head->prev_sibling_c = child;
+                }
+                else
+                        child->prev_sibling_c = child;
+
+                child->next_sibling = head;
+                node->first_child = child;
+        }
+
+        inline void insert_node_after(xml_node_struct* child, xml_node_struct* node)
+        {
+                xml_node_struct* parent = node->parent;
+
+                child->parent = parent;
+
+                if (node->next_sibling)
+                        node->next_sibling->prev_sibling_c = child;
+                else
+                        parent->first_child->prev_sibling_c = child;
+
+                child->next_sibling = node->next_sibling;
+                child->prev_sibling_c = node;
+
+                node->next_sibling = child;
+        }
+
+        inline void insert_node_before(xml_node_struct* child, xml_node_struct* node)
+        {
+                xml_node_struct* parent = node->parent;
+
+                child->parent = parent;
+
+                if (node->prev_sibling_c->next_sibling)
+                        node->prev_sibling_c->next_sibling = child;
+                else
+                        parent->first_child = child;
+
+                child->prev_sibling_c = node->prev_sibling_c;
+                child->next_sibling = node;
+
+                node->prev_sibling_c = child;
+        }
+
+        inline void remove_node(xml_node_struct* node)
+        {
+                xml_node_struct* parent = node->parent;
+
+                if (node->next_sibling)
+                        node->next_sibling->prev_sibling_c = node->prev_sibling_c;
+                else
+                        parent->first_child->prev_sibling_c = node->prev_sibling_c;
+
+                if (node->prev_sibling_c->next_sibling)
+                        node->prev_sibling_c->next_sibling = node->next_sibling;
+                else
+                        parent->first_child = node->next_sibling;
+
+                node->parent = 0;
+                node->prev_sibling_c = 0;
+                node->next_sibling = 0;
+        }
+
+        inline void append_attribute(xml_attribute_struct* attr, xml_node_struct* node)
+        {
+                xml_attribute_struct* head = node->first_attribute;
+
+                if (head)
+                {
+                        xml_attribute_struct* tail = head->prev_attribute_c;
+
+                        tail->next_attribute = attr;
+                        attr->prev_attribute_c = tail;
+                        head->prev_attribute_c = attr;
+                }
+                else
+                {
+                        node->first_attribute = attr;
+                        attr->prev_attribute_c = attr;
+                }
+        }
+
+        inline void prepend_attribute(xml_attribute_struct* attr, xml_node_struct* node)
+        {
+                xml_attribute_struct* head = node->first_attribute;
+
+                if (head)
+                {
+                        attr->prev_attribute_c = head->prev_attribute_c;
+                        head->prev_attribute_c = attr;
+                }
+                else
+                        attr->prev_attribute_c = attr;
+
+                attr->next_attribute = head;
+                node->first_attribute = attr;
+        }
+
+        inline void insert_attribute_after(xml_attribute_struct* attr, xml_attribute_struct* place, xml_node_struct* node)
+        {
+                if (place->next_attribute)
+                        place->next_attribute->prev_attribute_c = attr;
+                else
+                        node->first_attribute->prev_attribute_c = attr;
+
+                attr->next_attribute = place->next_attribute;
+                attr->prev_attribute_c = place;
+                place->next_attribute = attr;
+        }
+
+        inline void insert_attribute_before(xml_attribute_struct* attr, xml_attribute_struct* place, xml_node_struct* node)
+        {
+                if (place->prev_attribute_c->next_attribute)
+                        place->prev_attribute_c->next_attribute = attr;
+                else
+                        node->first_attribute = attr;
+
+                attr->prev_attribute_c = place->prev_attribute_c;
+                attr->next_attribute = place;
+                place->prev_attribute_c = attr;
+        }
+
+        inline void remove_attribute(xml_attribute_struct* attr, xml_node_struct* node)
+        {
+                if (attr->next_attribute)
+                        attr->next_attribute->prev_attribute_c = attr->prev_attribute_c;
+                else
+                        node->first_attribute->prev_attribute_c = attr->prev_attribute_c;
+
+                if (attr->prev_attribute_c->next_attribute)
+                        attr->prev_attribute_c->next_attribute = attr->next_attribute;
+                else
+                        node->first_attribute = attr->next_attribute;
+
+                attr->prev_attribute_c = 0;
+                attr->next_attribute = 0;
+        }
+
+        PUGI__FN_NO_INLINE xml_node_struct* append_new_node(xml_node_struct* node, xml_allocator& alloc, xml_node_type type = node_element)
+        {
+                if (!alloc.reserve()) return 0;
+
                 xml_node_struct* child = allocate_node(alloc, type);
                 if (!child) return 0;
 
-                child->parent = node;
-
-                xml_node_struct* first_child = node->first_child;
-                        
-                if (first_child)
-                {
-                        xml_node_struct* last_child = first_child->prev_sibling_c;
-
-                        last_child->next_sibling = child;
-                        child->prev_sibling_c = last_child;
-                        first_child->prev_sibling_c = child;
-                }
-                else
-                {
-                        node->first_child = child;
-                        child->prev_sibling_c = child;
-                }
-                        
+                append_node(child, node);
+
                 return child;
         }
 
-        PUGI__FN_NO_INLINE xml_attribute_struct* append_attribute_ll(xml_node_struct* node, xml_allocator& alloc)
-        {
-                xml_attribute_struct* a = allocate_attribute(alloc);
-                if (!a) return 0;
-
-                xml_attribute_struct* first_attribute = node->first_attribute;
-
-                if (first_attribute)
-                {
-                        xml_attribute_struct* last_attribute = first_attribute->prev_attribute_c;
-
-                        last_attribute->next_attribute = a;
-                        a->prev_attribute_c = last_attribute;
-                        first_attribute->prev_attribute_c = a;
-                }
-                else
-                {
-                        node->first_attribute = a;
-                        a->prev_attribute_c = a;
-                }
-                        
-                return a;
+        PUGI__FN_NO_INLINE xml_attribute_struct* append_new_attribute(xml_node_struct* node, xml_allocator& alloc)
+        {
+                if (!alloc.reserve()) return 0;
+
+                xml_attribute_struct* attr = allocate_attribute(alloc);
+                if (!attr) return 0;
+
+                append_attribute(attr, node);
+
+                return attr;
         }
 PUGI__NS_END
@@ -824 +1598 @@
         };
 
-        template <size_t size> struct wchar_selector;
-
-        template <> struct wchar_selector<2>
-        {
-                typedef uint16_t type;
-                typedef utf16_counter counter;
-                typedef utf16_writer writer;
-        };
-
-        template <> struct wchar_selector<4>
-        {
-                typedef uint32_t type;
-                typedef utf32_counter counter;
-                typedef utf32_writer writer;
-        };
-
-        typedef wchar_selector<sizeof(wchar_t)>::counter wchar_counter;
-        typedef wchar_selector<sizeof(wchar_t)>::writer wchar_writer;
-
-        template <typename Traits, typename opt_swap = opt_false> struct utf_decoder
-        {
-                static inline typename Traits::value_type decode_utf8_block(const uint8_t* data, size_t size, typename Traits::value_type result)
+        struct utf8_decoder
+        {
+                typedef uint8_t type;
+
+                template <typename Traits> static inline typename Traits::value_type process(const uint8_t* data, size_t size, typename Traits::value_type result, Traits)
                 {
                         const uint8_t utf8_byte_mask = 0x3f;
@@ -906 +1663 @@
                         return result;
                 }
-
-                static inline typename Traits::value_type decode_utf16_block(const uint16_t* data, size_t size, typename Traits::value_type result)
-                {
-                        const uint16_t* end = data + size;
-
-                        while (data < end)
+        };
+
+        template <typename opt_swap> struct utf16_decoder
+        {
+                typedef uint16_t type;
+
+                template <typename Traits> static inline typename Traits::value_type process(const uint16_t* data, size_t size, typename Traits::value_type result, Traits)
+                {
+                        while (size)
                         {
                                 uint16_t lead = opt_swap::value ? endian_swap(*data) : *data;
@@ -920 +1680 @@
                                         result = Traits::low(result, lead);
                                         data += 1;
+                                        size -= 1;
                                 }
                                 // U+E000..U+FFFF
@@ -926 +1687 @@
                                         result = Traits::low(result, lead);
                                         data += 1;
+                                        size -= 1;
                                 }
                                 // surrogate pair lead
-                                else if (static_cast<unsigned int>(lead - 0xD800) < 0x400 && data + 1 < end)
+                                else if (static_cast<unsigned int>(lead - 0xD800) < 0x400 && size >= 2)
                                 {
                                         uint16_t next = opt_swap::value ? endian_swap(data[1]) : data[1];
@@ -936 +1698 @@
                                                 result = Traits::high(result, 0x10000 + ((lead & 0x3ff) << 10) + (next & 0x3ff));
                                                 data += 2;
+                                                size -= 2;
                                         }
                                         else
                                         {
                                                 data += 1;
+                                                size -= 1;
                                         }
                                 }
@@ -945 +1709 @@
                                 {
                                         data += 1;
+                                        size -= 1;
                                 }
                         }
@@ -950 +1715 @@
                         return result;
                 }
-
-                static inline typename Traits::value_type decode_utf32_block(const uint32_t* data, size_t size, typename Traits::value_type result)
-                {
-                        const uint32_t* end = data + size;
-
-                        while (data < end)
+        };
+
+        template <typename opt_swap> struct utf32_decoder
+        {
+                typedef uint32_t type;
+
+                template <typename Traits> static inline typename Traits::value_type process(const uint32_t* data, size_t size, typename Traits::value_type result, Traits)
+                {
+                        while (size)
                         {
                                 uint32_t lead = opt_swap::value ? endian_swap(*data) : *data;
@@ -964 +1732 @@
                                         result = Traits::low(result, lead);
                                         data += 1;
+                                        size -= 1;
                                 }
                                 // U+10000..U+10FFFF
@@ -970 +1739 @@
                                         result = Traits::high(result, lead);
                                         data += 1;
+                                        size -= 1;
                                 }
                         }
@@ -975 +1745 @@
                         return result;
                 }
-
-                static inline typename Traits::value_type decode_latin1_block(const uint8_t* data, size_t size, typename Traits::value_type result)
-                {
-                        for (size_t i = 0; i < size; ++i)
-                        {
-                                result = Traits::low(result, data[i]);
+        };
+
+        struct latin1_decoder
+        {
+                typedef uint8_t type;
+
+                template <typename Traits> static inline typename Traits::value_type process(const uint8_t* data, size_t size, typename Traits::value_type result, Traits)
+                {
+                        while (size)
+                        {
+                                result = Traits::low(result, *data);
+                                data += 1;
+                                size -= 1;
                         }
 
                         return result;
                 }
-
-                static inline typename Traits::value_type decode_wchar_block_impl(const uint16_t* data, size_t size, typename Traits::value_type result)
-                {
-                        return decode_utf16_block(data, size, result);
-                }
-
-                static inline typename Traits::value_type decode_wchar_block_impl(const uint32_t* data, size_t size, typename Traits::value_type result)
-                {
-                        return decode_utf32_block(data, size, result);
-                }
-
-                static inline typename Traits::value_type decode_wchar_block(const wchar_t* data, size_t size, typename Traits::value_type result)
-                {
-                        return decode_wchar_block_impl(reinterpret_cast<const wchar_selector<sizeof(wchar_t)>::type*>(data), size, result);
-                }
         };
 
-        template <typename T> PUGI__FN void convert_utf_endian_swap(T* result, const T* data, size_t length)
-        {
-                for (size_t i = 0; i < length; ++i) result[i] = endian_swap(data[i]);
-        }
+        template <size_t size> struct wchar_selector;
+
+        template <> struct wchar_selector<2>
+        {
+                typedef uint16_t type;
+                typedef utf16_counter counter;
+                typedef utf16_writer writer;
+                typedef utf16_decoder<opt_false> decoder;
+        };
+
+        template <> struct wchar_selector<4>
+        {
+                typedef uint32_t type;
+                typedef utf32_counter counter;
+                typedef utf32_writer writer;
+                typedef utf32_decoder<opt_false> decoder;
+        };
+
+        typedef wchar_selector<sizeof(wchar_t)>::counter wchar_counter;
+        typedef wchar_selector<sizeof(wchar_t)>::writer wchar_writer;
+
+        struct wchar_decoder
+        {
+                typedef wchar_t type;
+
+                template <typename Traits> static inline typename Traits::value_type process(const wchar_t* data, size_t size, typename Traits::value_type result, Traits traits)
+                {
+                        typedef wchar_selector<sizeof(wchar_t)>::decoder decoder;
+
+                        return decoder::process(reinterpret_cast<const typename decoder::type*>(data), size, result, traits);
+                }
+        };
 
 #ifdef PUGIXML_WCHAR_MODE
         PUGI__FN void convert_wchar_endian_swap(wchar_t* result, const wchar_t* data, size_t length)
         {
-                for (size_t i = 0; i < length; ++i) result[i] = static_cast<wchar_t>(endian_swap(static_cast<wchar_selector<sizeof(wchar_t)>::type>(data[i])));
+                for (size_t i = 0; i < length; ++i)
+                        result[i] = static_cast<wchar_t>(endian_swap(static_cast<wchar_selector<sizeof(wchar_t)>::type>(data[i])));
         }
 #endif
@@ -1157 +1948 @@
         PUGI__FN bool get_mutable_buffer(char_t*& out_buffer, size_t& out_length, const void* contents, size_t size, bool is_mutable)
         {
+                size_t length = size / sizeof(char_t);
+
                 if (is_mutable)
                 {
                         out_buffer = static_cast<char_t*>(const_cast<void*>(contents));
+                        out_length = length;
                 }
                 else
                 {
-                        void* buffer = xml_memory::allocate(size > 0 ? size : 1);
+                        char_t* buffer = static_cast<char_t*>(xml_memory::allocate((length + 1) * sizeof(char_t)));
                         if (!buffer) return false;
 
-                        memcpy(buffer, contents, size);
-
-                        out_buffer = static_cast<char_t*>(buffer);
-                }
-
-                out_length = size / sizeof(char_t);
+                        if (contents)
+                                memcpy(buffer, contents, length * sizeof(char_t));
+                        else
+                                assert(length == 0);
+
+                        buffer[length] = 0;
+
+                        out_buffer = buffer;
+                        out_length = length + 1;
+                }
 
                 return true;
@@ -1186 +1984 @@
         {
                 const char_t* data = static_cast<const char_t*>(contents);
-        
+                size_t length = size / sizeof(char_t);
+
                 if (is_mutable)
                 {
-                        out_buffer = const_cast<char_t*>(data);
+                        char_t* buffer = const_cast<char_t*>(data);
+
+                        convert_wchar_endian_swap(buffer, data, length);
+
+                        out_buffer = buffer;
+                        out_length = length;
                 }
                 else
                 {
-                        out_buffer = static_cast<char_t*>(xml_memory::allocate(size > 0 ? size : 1));
-                        if (!out_buffer) return false;
-                }
-
-                out_length = size / sizeof(char_t);
-
-                convert_wchar_endian_swap(out_buffer, data, out_length);
+                        char_t* buffer = static_cast<char_t*>(xml_memory::allocate((length + 1) * sizeof(char_t)));
+                        if (!buffer) return false;
+
+                        convert_wchar_endian_swap(buffer, data, length);
+                        buffer[length] = 0;
+
+                        out_buffer = buffer;
+                        out_length = length + 1;
+                }
 
                 return true;
         }
 
-        PUGI__FN bool convert_buffer_utf8(char_t*& out_buffer, size_t& out_length, const void* contents, size_t size)
-        {
-                const uint8_t* data = static_cast<const uint8_t*>(contents);
+        template <typename D> PUGI__FN bool convert_buffer_generic(char_t*& out_buffer, size_t& out_length, const void* contents, size_t size, D)
+        {
+                const typename D::type* data = static_cast<const typename D::type*>(contents);
+                size_t data_length = size / sizeof(typename D::type);
 
                 // first pass: get length in wchar_t units
-                out_length = utf_decoder<wchar_counter>::decode_utf8_block(data, size, 0);
+                size_t length = D::process(data, data_length, 0, wchar_counter());
 
                 // allocate buffer of suitable length
-                out_buffer = static_cast<char_t*>(xml_memory::allocate((out_length > 0 ? out_length : 1) * sizeof(char_t)));
-                if (!out_buffer) return false;
-
-                // second pass: convert utf8 input to wchar_t
-                wchar_writer::value_type out_begin = reinterpret_cast<wchar_writer::value_type>(out_buffer);
-                wchar_writer::value_type out_end = utf_decoder<wchar_writer>::decode_utf8_block(data, size, out_begin);
-
-                assert(out_end == out_begin + out_length);
-                (void)!out_end;
-
-                return true;
-        }
-
-        template <typename opt_swap> PUGI__FN bool convert_buffer_utf16(char_t*& out_buffer, size_t& out_length, const void* contents, size_t size, opt_swap)
-        {
-                const uint16_t* data = static_cast<const uint16_t*>(contents);
-                size_t length = size / sizeof(uint16_t);
-
-                // first pass: get length in wchar_t units
-                out_length = utf_decoder<wchar_counter, opt_swap>::decode_utf16_block(data, length, 0);
-
-                // allocate buffer of suitable length
-                out_buffer = static_cast<char_t*>(xml_memory::allocate((out_length > 0 ? out_length : 1) * sizeof(char_t)));
-                if (!out_buffer) return false;
+                char_t* buffer = static_cast<char_t*>(xml_memory::allocate((length + 1) * sizeof(char_t)));
+                if (!buffer) return false;
 
                 // second pass: convert utf16 input to wchar_t
-                wchar_writer::value_type out_begin = reinterpret_cast<wchar_writer::value_type>(out_buffer);
-                wchar_writer::value_type out_end = utf_decoder<wchar_writer, opt_swap>::decode_utf16_block(data, length, out_begin);
-
-                assert(out_end == out_begin + out_length);
-                (void)!out_end;
-
-                return true;
-        }
-
-        template <typename opt_swap> PUGI__FN bool convert_buffer_utf32(char_t*& out_buffer, size_t& out_length, const void* contents, size_t size, opt_swap)
-        {
-                const uint32_t* data = static_cast<const uint32_t*>(contents);
-                size_t length = size / sizeof(uint32_t);
-
-                // first pass: get length in wchar_t units
-                out_length = utf_decoder<wchar_counter, opt_swap>::decode_utf32_block(data, length, 0);
-
-                // allocate buffer of suitable length
-                out_buffer = static_cast<char_t*>(xml_memory::allocate((out_length > 0 ? out_length : 1) * sizeof(char_t)));
-                if (!out_buffer) return false;
-
-                // second pass: convert utf32 input to wchar_t
-                wchar_writer::value_type out_begin = reinterpret_cast<wchar_writer::value_type>(out_buffer);
-                wchar_writer::value_type out_end = utf_decoder<wchar_writer, opt_swap>::decode_utf32_block(data, length, out_begin);
-
-                assert(out_end == out_begin + out_length);
-                (void)!out_end;
-
-                return true;
-        }
-
-        PUGI__FN bool convert_buffer_latin1(char_t*& out_buffer, size_t& out_length, const void* contents, size_t size)
-        {
-                const uint8_t* data = static_cast<const uint8_t*>(contents);
-
-                // get length in wchar_t units
-                out_length = size;
-
-                // allocate buffer of suitable length
-                out_buffer = static_cast<char_t*>(xml_memory::allocate((out_length > 0 ? out_length : 1) * sizeof(char_t)));
-                if (!out_buffer) return false;
-
-                // convert latin1 input to wchar_t
-                wchar_writer::value_type out_begin = reinterpret_cast<wchar_writer::value_type>(out_buffer);
-                wchar_writer::value_type out_end = utf_decoder<wchar_writer>::decode_latin1_block(data, size, out_begin);
-
-                assert(out_end == out_begin + out_length);
-                (void)!out_end;
+                wchar_writer::value_type obegin = reinterpret_cast<wchar_writer::value_type>(buffer);
+                wchar_writer::value_type oend = D::process(data, data_length, obegin, wchar_writer());
+
+                assert(oend == obegin + length);
+                *oend = 0;
+
+                out_buffer = buffer;
+                out_length = length + 1;
 
                 return true;
@@ -1296 +2041 @@
 
                 // fast path: no conversion required
-                if (encoding == wchar_encoding) return get_mutable_buffer(out_buffer, out_length, contents, size, is_mutable);
+                if (encoding == wchar_encoding)
+                        return get_mutable_buffer(out_buffer, out_length, contents, size, is_mutable);
 
                 // only endian-swapping is required
-                if (need_endian_swap_utf(encoding, wchar_encoding)) return convert_buffer_endian_swap(out_buffer, out_length, contents, size, is_mutable);
+                if (need_endian_swap_utf(encoding, wchar_encoding))
+                        return convert_buffer_endian_swap(out_buffer, out_length, contents, size, is_mutable);
 
                 // source encoding is utf8
-                if (encoding == encoding_utf8) return convert_buffer_utf8(out_buffer, out_length, contents, size);
+                if (encoding == encoding_utf8)
+                        return convert_buffer_generic(out_buffer, out_length, contents, size, utf8_decoder());
 
                 // source encoding is utf16
@@ -1310 +2058 @@
 
                         return (native_encoding == encoding) ?
-                                convert_buffer_utf16(out_buffer, out_length, contents, size, opt_false()) :
-                                convert_buffer_utf16(out_buffer, out_length, contents, size, opt_true());
+                                convert_buffer_generic(out_buffer, out_length, contents, size, utf16_decoder<opt_false>()) :
+                                convert_buffer_generic(out_buffer, out_length, contents, size, utf16_decoder<opt_true>());
                 }
 
@@ -1320 +2068 @@
 
                         return (native_encoding == encoding) ?
-                                convert_buffer_utf32(out_buffer, out_length, contents, size, opt_false()) :
-                                convert_buffer_utf32(out_buffer, out_length, contents, size, opt_true());
+                                convert_buffer_generic(out_buffer, out_length, contents, size, utf32_decoder<opt_false>()) :
+                                convert_buffer_generic(out_buffer, out_length, contents, size, utf32_decoder<opt_true>());
                 }
 
                 // source encoding is latin1
-                if (encoding == encoding_latin1) return convert_buffer_latin1(out_buffer, out_length, contents, size);
+                if (encoding == encoding_latin1)
+                        return convert_buffer_generic(out_buffer, out_length, contents, size, latin1_decoder());
 
                 assert(!"Invalid encoding");
@@ -1331 +2080 @@
         }
 #else
-        template <typename opt_swap> PUGI__FN bool convert_buffer_utf16(char_t*& out_buffer, size_t& out_length, const void* contents, size_t size, opt_swap)
-        {
-                const uint16_t* data = static_cast<const uint16_t*>(contents);
-                size_t length = size / sizeof(uint16_t);
+        template <typename D> PUGI__FN bool convert_buffer_generic(char_t*& out_buffer, size_t& out_length, const void* contents, size_t size, D)
+        {
+                const typename D::type* data = static_cast<const typename D::type*>(contents);
+                size_t data_length = size / sizeof(typename D::type);
 
                 // first pass: get length in utf8 units
-                out_length = utf_decoder<utf8_counter, opt_swap>::decode_utf16_block(data, length, 0);
+                size_t length = D::process(data, data_length, 0, utf8_counter());
 
                 // allocate buffer of suitable length
-                out_buffer = static_cast<char_t*>(xml_memory::allocate((out_length > 0 ? out_length : 1) * sizeof(char_t)));
-                if (!out_buffer) return false;
+                char_t* buffer = static_cast<char_t*>(xml_memory::allocate((length + 1) * sizeof(char_t)));
+                if (!buffer) return false;
 
                 // second pass: convert utf16 input to utf8
-                uint8_t* out_begin = reinterpret_cast<uint8_t*>(out_buffer);
-                uint8_t* out_end = utf_decoder<utf8_writer, opt_swap>::decode_utf16_block(data, length, out_begin);
-
-                assert(out_end == out_begin + out_length);
-                (void)!out_end;
-
-                return true;
-        }
-
-        template <typename opt_swap> PUGI__FN bool convert_buffer_utf32(char_t*& out_buffer, size_t& out_length, const void* contents, size_t size, opt_swap)
-        {
-                const uint32_t* data = static_cast<const uint32_t*>(contents);
-                size_t length = size / sizeof(uint32_t);
-
-                // first pass: get length in utf8 units
-                out_length = utf_decoder<utf8_counter, opt_swap>::decode_utf32_block(data, length, 0);
-
-                // allocate buffer of suitable length
-                out_buffer = static_cast<char_t*>(xml_memory::allocate((out_length > 0 ? out_length : 1) * sizeof(char_t)));
-                if (!out_buffer) return false;
-
-                // second pass: convert utf32 input to utf8
-                uint8_t* out_begin = reinterpret_cast<uint8_t*>(out_buffer);
-                uint8_t* out_end = utf_decoder<utf8_writer, opt_swap>::decode_utf32_block(data, length, out_begin);
-
-                assert(out_end == out_begin + out_length);
-                (void)!out_end;
+                uint8_t* obegin = reinterpret_cast<uint8_t*>(buffer);
+                uint8_t* oend = D::process(data, data_length, obegin, utf8_writer());
+
+                assert(oend == obegin + length);
+                *oend = 0;
+
+                out_buffer = buffer;
+                out_length = length + 1;
 
                 return true;
@@ -1387 +2117 @@
         {
                 const uint8_t* data = static_cast<const uint8_t*>(contents);
+                size_t data_length = size;
 
                 // get size of prefix that does not need utf8 conversion
-                size_t prefix_length = get_latin1_7bit_prefix_length(data, size);
-                assert(prefix_length <= size);
+                size_t prefix_length = get_latin1_7bit_prefix_length(data, data_length);
+                assert(prefix_length <= data_length);
 
                 const uint8_t* postfix = data + prefix_length;
-                size_t postfix_length = size - prefix_length;
+                size_t postfix_length = data_length - prefix_length;
 
                 // if no conversion is needed, just return the original buffer
@@ -1399 +2130 @@
 
                 // first pass: get length in utf8 units
-                out_length = prefix_length + utf_decoder<utf8_counter>::decode_latin1_block(postfix, postfix_length, 0);
+                size_t length = prefix_length + latin1_decoder::process(postfix, postfix_length, 0, utf8_counter());
 
                 // allocate buffer of suitable length
-                out_buffer = static_cast<char_t*>(xml_memory::allocate((out_length > 0 ? out_length : 1) * sizeof(char_t)));
-                if (!out_buffer) return false;
+                char_t* buffer = static_cast<char_t*>(xml_memory::allocate((length + 1) * sizeof(char_t)));
+                if (!buffer) return false;
 
                 // second pass: convert latin1 input to utf8
-                memcpy(out_buffer, data, prefix_length);
-
-                uint8_t* out_begin = reinterpret_cast<uint8_t*>(out_buffer);
-                uint8_t* out_end = utf_decoder<utf8_writer>::decode_latin1_block(postfix, postfix_length, out_begin + prefix_length);
-
-                assert(out_end == out_begin + out_length);
-                (void)!out_end;
+                memcpy(buffer, data, prefix_length);
+
+                uint8_t* obegin = reinterpret_cast<uint8_t*>(buffer);
+                uint8_t* oend = latin1_decoder::process(postfix, postfix_length, obegin + prefix_length, utf8_writer());
+
+                assert(oend == obegin + length);
+                *oend = 0;
+
+                out_buffer = buffer;
+                out_length = length + 1;
 
                 return true;
@@ -1420 +2154 @@
         {
                 // fast path: no conversion required
-                if (encoding == encoding_utf8) return get_mutable_buffer(out_buffer, out_length, contents, size, is_mutable);
+                if (encoding == encoding_utf8)
+                        return get_mutable_buffer(out_buffer, out_length, contents, size, is_mutable);
 
                 // source encoding is utf16
@@ -1428 +2163 @@
 
                         return (native_encoding == encoding) ?
-                                convert_buffer_utf16(out_buffer, out_length, contents, size, opt_false()) :
-                                convert_buffer_utf16(out_buffer, out_length, contents, size, opt_true());
+                                convert_buffer_generic(out_buffer, out_length, contents, size, utf16_decoder<opt_false>()) :
+                                convert_buffer_generic(out_buffer, out_length, contents, size, utf16_decoder<opt_true>());
                 }
 
@@ -1438 +2173 @@
 
                         return (native_encoding == encoding) ?
-                                convert_buffer_utf32(out_buffer, out_length, contents, size, opt_false()) :
-                                convert_buffer_utf32(out_buffer, out_length, contents, size, opt_true());
+                                convert_buffer_generic(out_buffer, out_length, contents, size, utf32_decoder<opt_false>()) :
+                                convert_buffer_generic(out_buffer, out_length, contents, size, utf32_decoder<opt_true>());
                 }
 
                 // source encoding is latin1
-                if (encoding == encoding_latin1) return convert_buffer_latin1(out_buffer, out_length, contents, size, is_mutable);
+                if (encoding == encoding_latin1)
+                        return convert_buffer_latin1(out_buffer, out_length, contents, size, is_mutable);
 
                 assert(!"Invalid encoding");
@@ -1453 +2189 @@
         {
                 // get length in utf8 characters
-                return utf_decoder<utf8_counter>::decode_wchar_block(str, length, 0);
+                return wchar_decoder::process(str, length, 0, utf8_counter());
         }
 
@@ -1460 +2196 @@
                 // convert to utf8
                 uint8_t* begin = reinterpret_cast<uint8_t*>(buffer);
-                uint8_t* end = utf_decoder<utf8_writer>::decode_wchar_block(str, length, begin);
+                uint8_t* end = wchar_decoder::process(str, length, begin, utf8_writer());
         
                 assert(begin + size == end);
                 (void)!end;
-
-                // zero-terminate
-                buffer[size] = 0;
+                (void)!size;
         }
         
@@ -1490 +2224 @@
 
                 // first pass: get length in wchar_t units
-                size_t length = utf_decoder<wchar_counter>::decode_utf8_block(data, size, 0);
+                size_t length = utf8_decoder::process(data, size, 0, wchar_counter());
 
                 // allocate resulting string
@@ -1500 +2234 @@
                 {
                         wchar_writer::value_type begin = reinterpret_cast<wchar_writer::value_type>(&result[0]);
-                        wchar_writer::value_type end = utf_decoder<wchar_writer>::decode_utf8_block(data, size, begin);
+                        wchar_writer::value_type end = utf8_decoder::process(data, size, begin, wchar_writer());
 
                         assert(begin + length == end);
@@ -1510 +2244 @@
 #endif
 
-        inline bool strcpy_insitu_allow(size_t length, uintptr_t allocated, char_t* target)
-        {
-                assert(target);
+        template <typename Header>
+        inline bool strcpy_insitu_allow(size_t length, const Header& header, uintptr_t header_mask, char_t* target)
+        {
+                // never reuse shared memory
+                if (header & xml_memory_page_contents_shared_mask) return false;
+
                 size_t target_length = strlength(target);
 
                 // always reuse document buffer memory if possible
-                if (!allocated) return target_length >= length;
+                if ((header & header_mask) == 0) return target_length >= length;
 
                 // reuse heap memory if waste is not too great
@@ -1524 +2261 @@
         }
 
-        PUGI__FN bool strcpy_insitu(char_t*& dest, uintptr_t& header, uintptr_t header_mask, const char_t* source)
-        {
-                size_t source_length = strlength(source);
-
+        template <typename String, typename Header>
+        PUGI__FN bool strcpy_insitu(String& dest, Header& header, uintptr_t header_mask, const char_t* source, size_t source_length)
+        {
                 if (source_length == 0)
                 {
                         // empty string and null pointer are equivalent, so just deallocate old memory
-                        xml_allocator* alloc = reinterpret_cast<xml_memory_page*>(header & xml_memory_page_pointer_mask)->allocator;
+                        xml_allocator* alloc = PUGI__GETPAGE_IMPL(header)->allocator;
 
                         if (header & header_mask) alloc->deallocate_string(dest);
@@ -1541 +2277 @@
                         return true;
                 }
-                else if (dest && strcpy_insitu_allow(source_length, header & header_mask, dest))
+                else if (dest && strcpy_insitu_allow(source_length, header, header_mask, dest))
                 {
                         // we can reuse old buffer, so just copy the new data (including zero terminator)
-                        memcpy(dest, source, (source_length + 1) * sizeof(char_t));
+                        memcpy(dest, source, source_length * sizeof(char_t));
+                        dest[source_length] = 0;
                         
                         return true;
@@ -1550 +2287 @@
                 else
                 {
-                        xml_allocator* alloc = reinterpret_cast<xml_memory_page*>(header & xml_memory_page_pointer_mask)->allocator;
+                        xml_allocator* alloc = PUGI__GETPAGE_IMPL(header)->allocator;
+
+                        if (!alloc->reserve()) return false;
 
                         // allocate new buffer
@@ -1557 +2296 @@
 
                         // copy the string (including zero terminator)
-                        memcpy(buf, source, (source_length + 1) * sizeof(char_t));
+                        memcpy(buf, source, source_length * sizeof(char_t));
+                        buf[source_length] = 0;
 
                         // deallocate old buffer (*after* the above to protect against overlapping memory and/or allocation failures)
@@ -1654 +2394 @@
                                         for (;;)
                                         {
-                                                if (static_cast<unsigned int>(ch - '0') <= 9)
+                                                if (static_cast<unsigned int>(static_cast<unsigned int>(ch) - '0') <= 9)
                                                         ucsc = 10 * ucsc + (ch - '0');
                                                 else if (ch == ';')
@@ -1752 +2492 @@
         }
 
-        // Utility macro for last character handling
-        #define ENDSWITH(c, e) ((c) == (e) || ((c) == 0 && endch == (e)))
+        // Parser utilities
+        #define PUGI__ENDSWITH(c, e)        ((c) == (e) || ((c) == 0 && endch == (e)))
+        #define PUGI__SKIPWS()              { while (PUGI__IS_CHARTYPE(*s, ct_space)) ++s; }
+        #define PUGI__OPTSET(OPT)           ( optmsk & (OPT) )
+        #define PUGI__PUSHNODE(TYPE)        { cursor = append_new_node(cursor, alloc, TYPE); if (!cursor) PUGI__THROW_ERROR(status_out_of_memory, s); }
+        #define PUGI__POPNODE()             { cursor = cursor->parent; }
+        #define PUGI__SCANFOR(X)            { while (*s != 0 && !(X)) ++s; }
+        #define PUGI__SCANWHILE(X)          { while (X) ++s; }
+        #define PUGI__SCANWHILE_UNROLL(X)   { for (;;) { char_t ss = s[0]; if (PUGI__UNLIKELY(!(X))) { break; } ss = s[1]; if (PUGI__UNLIKELY(!(X))) { s += 1; break; } ss = s[2]; if (PUGI__UNLIKELY(!(X))) { s += 2; break; } ss = s[3]; if (PUGI__UNLIKELY(!(X))) { s += 3; break; } s += 4; } }
+        #define PUGI__ENDSEG()              { ch = *s; *s = 0; ++s; }
+        #define PUGI__THROW_ERROR(err, m)   return error_offset = m, error_status = err, static_cast<char_t*>(0)
+        #define PUGI__CHECK_ERROR(err, m)   { if (*s == 0) PUGI__THROW_ERROR(err, m); }
 
         PUGI__FN char_t* strconv_comment(char_t* s, char_t endch)
@@ -1761 +2511 @@
                 while (true)
                 {
-                        while (!PUGI__IS_CHARTYPE(*s, ct_parse_comment)) ++s;
+                        PUGI__SCANWHILE_UNROLL(!PUGI__IS_CHARTYPE(ss, ct_parse_comment));
                 
                         if (*s == '\r') // Either a single 0x0d or 0x0d 0x0a pair
@@ -1769 +2519 @@
                                 if (*s == '\n') g.push(s, 1);
                         }
-                        else if (s[0] == '-' && s[1] == '-' && ENDSWITH(s[2], '>')) // comment ends here
+                        else if (s[0] == '-' && s[1] == '-' && PUGI__ENDSWITH(s[2], '>')) // comment ends here
                         {
                                 *g.flush(s) = 0;
@@ -1789 +2539 @@
                 while (true)
                 {
-                        while (!PUGI__IS_CHARTYPE(*s, ct_parse_cdata)) ++s;
+                        PUGI__SCANWHILE_UNROLL(!PUGI__IS_CHARTYPE(ss, ct_parse_cdata));
                         
                         if (*s == '\r') // Either a single 0x0d or 0x0d 0x0a pair
@@ -1797 +2547 @@
                                 if (*s == '\n') g.push(s, 1);
                         }
-                        else if (s[0] == ']' && s[1] == ']' && ENDSWITH(s[2], '>')) // CDATA ends here
+                        else if (s[0] == ']' && s[1] == ']' && PUGI__ENDSWITH(s[2], '>')) // CDATA ends here
                         {
                                 *g.flush(s) = 0;
@@ -1813 +2563 @@
         typedef char_t* (*strconv_pcdata_t)(char_t*);
                 
-        template <typename opt_eol, typename opt_escape> struct strconv_pcdata_impl
+        template <typename opt_trim, typename opt_eol, typename opt_escape> struct strconv_pcdata_impl
         {
                 static char_t* parse(char_t* s)
                 {
                         gap g;
-                        
+
+                        char_t* begin = s;
+
                         while (true)
                         {
-                                while (!PUGI__IS_CHARTYPE(*s, ct_parse_pcdata)) ++s;
-                                        
+                                PUGI__SCANWHILE_UNROLL(!PUGI__IS_CHARTYPE(ss, ct_parse_pcdata));
+
                                 if (*s == '<') // PCDATA ends here
                                 {
-                                        *g.flush(s) = 0;
+                                        char_t* end = g.flush(s);
+
+                                        if (opt_trim::value)
+                                                while (end > begin && PUGI__IS_CHARTYPE(end[-1], ct_space))
+                                                        --end;
+
+                                        *end = 0;
                                         
                                         return s + 1;
@@ -1841 +2599 @@
                                 else if (*s == 0)
                                 {
+                                        char_t* end = g.flush(s);
+
+                                        if (opt_trim::value)
+                                                while (end > begin && PUGI__IS_CHARTYPE(end[-1], ct_space))
+                                                        --end;
+
+                                        *end = 0;
+
                                         return s;
                                 }
@@ -1850 +2616 @@
         PUGI__FN strconv_pcdata_t get_strconv_pcdata(unsigned int optmask)
         {
-                PUGI__STATIC_ASSERT(parse_escapes == 0x10 && parse_eol == 0x20);
-
-                switch ((optmask >> 4) & 3) // get bitmask for flags (eol escapes)
-                {
-                case 0: return strconv_pcdata_impl<opt_false, opt_false>::parse;
-                case 1: return strconv_pcdata_impl<opt_false, opt_true>::parse;
-                case 2: return strconv_pcdata_impl<opt_true, opt_false>::parse;
-                case 3: return strconv_pcdata_impl<opt_true, opt_true>::parse;
-                default: return 0; // should not get here
+                PUGI__STATIC_ASSERT(parse_escapes == 0x10 && parse_eol == 0x20 && parse_trim_pcdata == 0x0800);
+
+                switch (((optmask >> 4) & 3) | ((optmask >> 9) & 4)) // get bitmask for flags (eol escapes trim)
+                {
+                case 0: return strconv_pcdata_impl<opt_false, opt_false, opt_false>::parse;
+                case 1: return strconv_pcdata_impl<opt_false, opt_false, opt_true>::parse;
+                case 2: return strconv_pcdata_impl<opt_false, opt_true, opt_false>::parse;
+                case 3: return strconv_pcdata_impl<opt_false, opt_true, opt_true>::parse;
+                case 4: return strconv_pcdata_impl<opt_true, opt_false, opt_false>::parse;
+                case 5: return strconv_pcdata_impl<opt_true, opt_false, opt_true>::parse;
+                case 6: return strconv_pcdata_impl<opt_true, opt_true, opt_false>::parse;
+                case 7: return strconv_pcdata_impl<opt_true, opt_true, opt_true>::parse;
+                default: assert(false); return 0; // should not get here
                 }
         }
@@ -1883 +2653 @@
                         while (true)
                         {
-                                while (!PUGI__IS_CHARTYPE(*s, ct_parse_attr_ws | ct_space)) ++s;
+                                PUGI__SCANWHILE_UNROLL(!PUGI__IS_CHARTYPE(ss, ct_parse_attr_ws | ct_space));
                                 
                                 if (*s == end_quote)
@@ -1924 +2694 @@
                         while (true)
                         {
-                                while (!PUGI__IS_CHARTYPE(*s, ct_parse_attr_ws)) ++s;
+                                PUGI__SCANWHILE_UNROLL(!PUGI__IS_CHARTYPE(ss, ct_parse_attr_ws));
                                 
                                 if (*s == end_quote)
@@ -1960 +2730 @@
                         while (true)
                         {
-                                while (!PUGI__IS_CHARTYPE(*s, ct_parse_attr)) ++s;
+                                PUGI__SCANWHILE_UNROLL(!PUGI__IS_CHARTYPE(ss, ct_parse_attr));
                                 
                                 if (*s == end_quote)
@@ -1992 +2762 @@
                         while (true)
                         {
-                                while (!PUGI__IS_CHARTYPE(*s, ct_parse_attr)) ++s;
+                                PUGI__SCANWHILE_UNROLL(!PUGI__IS_CHARTYPE(ss, ct_parse_attr));
                                 
                                 if (*s == end_quote)
@@ -2035 +2805 @@
                 case 14: return strconv_attribute_impl<opt_false>::parse_wnorm;
                 case 15: return strconv_attribute_impl<opt_true>::parse_wnorm;
-                default: return 0; // should not get here
+                default: assert(false); return 0; // should not get here
                 }
         }
@@ -2051 +2821 @@
         {
                 xml_allocator alloc;
+                xml_allocator* alloc_state;
                 char_t* error_offset;
                 xml_parse_status error_status;
                 
-                // Parser utilities.
-                #define PUGI__SKIPWS()                  { while (PUGI__IS_CHARTYPE(*s, ct_space)) ++s; }
-                #define PUGI__OPTSET(OPT)                       ( optmsk & (OPT) )
-                #define PUGI__PUSHNODE(TYPE)            { cursor = append_node(cursor, alloc, TYPE); if (!cursor) PUGI__THROW_ERROR(status_out_of_memory, s); }
-                #define PUGI__POPNODE()                 { cursor = cursor->parent; }
-                #define PUGI__SCANFOR(X)                        { while (*s != 0 && !(X)) ++s; }
-                #define PUGI__SCANWHILE(X)              { while ((X)) ++s; }
-                #define PUGI__ENDSEG()                  { ch = *s; *s = 0; ++s; }
-                #define PUGI__THROW_ERROR(err, m)       return error_offset = m, error_status = err, static_cast<char_t*>(0)
-                #define PUGI__CHECK_ERROR(err, m)       { if (*s == 0) PUGI__THROW_ERROR(err, m); }
-                
-                xml_parser(const xml_allocator& alloc_): alloc(alloc_), error_offset(0), error_status(status_ok)
-                {
+                xml_parser(xml_allocator* alloc_): alloc(*alloc_), alloc_state(alloc_), error_offset(0), error_status(status_ok)
+                {
+                }
+
+                ~xml_parser()
+                {
+                        *alloc_state = alloc;
                 }
 
@@ -2102 +2867 @@
                                 if (!*s) PUGI__THROW_ERROR(status_bad_doctype, s);
 
-                                s += 4;
+                                s += 3;
                         }
                         else PUGI__THROW_ERROR(status_bad_doctype, s);
@@ -2111 +2876 @@
                 char_t* parse_doctype_ignore(char_t* s)
                 {
+                        size_t depth = 0;
+
                         assert(s[0] == '<' && s[1] == '!' && s[2] == '[');
-                        s++;
+                        s += 3;
 
                         while (*s)
@@ -2119 +2886 @@
                                 {
                                         // nested ignore section
-                                        s = parse_doctype_ignore(s);
-                                        if (!s) return s;
+                                        s += 3;
+                                        depth++;
                                 }
                                 else if (s[0] == ']' && s[1] == ']' && s[2] == '>')
@@ -2127 +2894 @@
                                         s += 3;
 
-                                        return s;
+                                        if (depth == 0)
+                                                return s;
+
+                                        depth--;
                                 }
                                 else s++;
@@ -2135 +2905 @@
                 }
 
-                char_t* parse_doctype_group(char_t* s, char_t endch, bool toplevel)
-                {
-                        assert(s[0] == '<' && s[1] == '!');
-                        s++;
+                char_t* parse_doctype_group(char_t* s, char_t endch)
+                {
+                        size_t depth = 0;
+
+                        assert((s[0] == '<' || s[0] == 0) && s[1] == '!');
+                        s += 2;
 
                         while (*s)
@@ -2153 +2925 @@
                                         {
                                                 // some control group
-                                                s = parse_doctype_group(s, endch, false);
-                                                if (!s) return s;
+                                                s += 2;
+                                                depth++;
                                         }
                                 }
@@ -2165 +2937 @@
                                 else if (*s == '>')
                                 {
+                                        if (depth == 0)
+                                                return s;
+
+                                        depth--;
                                         s++;
-
-                                        return s;
                                 }
                                 else s++;
                         }
 
-                        if (!toplevel || endch != '>') PUGI__THROW_ERROR(status_bad_doctype, s);
+                        if (depth != 0 || endch != '>') PUGI__THROW_ERROR(status_bad_doctype, s);
 
                         return s;
@@ -2205 +2979 @@
                                         {
                                                 // Scan for terminating '-->'.
-                                                PUGI__SCANFOR(s[0] == '-' && s[1] == '-' && ENDSWITH(s[2], '>'));
+                                                PUGI__SCANFOR(s[0] == '-' && s[1] == '-' && PUGI__ENDSWITH(s[2], '>'));
                                                 PUGI__CHECK_ERROR(status_bad_comment, s);
 
@@ -2237 +3011 @@
                                                 {
                                                         // Scan for terminating ']]>'.
-                                                        PUGI__SCANFOR(s[0] == ']' && s[1] == ']' && ENDSWITH(s[2], '>'));
+                                                        PUGI__SCANFOR(s[0] == ']' && s[1] == ']' && PUGI__ENDSWITH(s[2], '>'));
                                                         PUGI__CHECK_ERROR(status_bad_cdata, s);
 
@@ -2246 +3020 @@
                                         {
                                                 // Scan for terminating ']]>'.
-                                                PUGI__SCANFOR(s[0] == ']' && s[1] == ']' && ENDSWITH(s[2], '>'));
+                                                PUGI__SCANFOR(s[0] == ']' && s[1] == ']' && PUGI__ENDSWITH(s[2], '>'));
                                                 PUGI__CHECK_ERROR(status_bad_cdata, s);
 
@@ -2256 +3030 @@
                                 else PUGI__THROW_ERROR(status_bad_cdata, s);
                         }
-                        else if (s[0] == 'D' && s[1] == 'O' && s[2] == 'C' && s[3] == 'T' && s[4] == 'Y' && s[5] == 'P' && ENDSWITH(s[6], 'E'))
+                        else if (s[0] == 'D' && s[1] == 'O' && s[2] == 'C' && s[3] == 'T' && s[4] == 'Y' && s[5] == 'P' && PUGI__ENDSWITH(s[6], 'E'))
                         {
                                 s -= 2;
@@ -2264 +3038 @@
                                 char_t* mark = s + 9;
 
-                                s = parse_doctype_group(s, endch, true);
+                                s = parse_doctype_group(s, endch);
                                 if (!s) return s;
 
+                                assert((*s == 0 && endch == '>') || *s == '>');
+                                if (*s) *s++ = 0;
+
                                 if (PUGI__OPTSET(parse_doctype))
                                 {
@@ -2274 +3051 @@
 
                                         cursor->value = mark;
-
-                                        assert((s[0] == 0 && endch == '>') || s[-1] == '>');
-                                        s[*s == 0 ? 0 : -1] = 0;
-
-                                        PUGI__POPNODE();
                                 }
                         }
@@ -2330 +3102 @@
                                 {
                                         // empty node
-                                        if (!ENDSWITH(*s, '>')) PUGI__THROW_ERROR(status_bad_pi, s);
+                                        if (!PUGI__ENDSWITH(*s, '>')) PUGI__THROW_ERROR(status_bad_pi, s);
                                         s += (*s == '>');
 
@@ -2342 +3114 @@
                                         char_t* value = s;
 
-                                        PUGI__SCANFOR(s[0] == '?' && ENDSWITH(s[1], '>'));
+                                        PUGI__SCANFOR(s[0] == '?' && PUGI__ENDSWITH(s[1], '>'));
                                         PUGI__CHECK_ERROR(status_bad_pi, s);
 
@@ -2357 +3129 @@
                                                 // store value and step over >
                                                 cursor->value = value;
+
                                                 PUGI__POPNODE();
 
@@ -2369 +3142 @@
                         {
                                 // scan for tag end
-                                PUGI__SCANFOR(s[0] == '?' && ENDSWITH(s[1], '>'));
+                                PUGI__SCANFOR(s[0] == '?' && PUGI__ENDSWITH(s[1], '>'));
                                 PUGI__CHECK_ERROR(status_bad_pi, s);
 
@@ -2381 +3154 @@
                 }
 
-                char_t* parse(char_t* s, xml_node_struct* xmldoc, unsigned int optmsk, char_t endch)
+                char_t* parse_tree(char_t* s, xml_node_struct* root, unsigned int optmsk, char_t endch)
                 {
                         strconv_attribute_t strconv_attribute = get_strconv_attribute(optmsk);
@@ -2387 +3160 @@
                         
                         char_t ch = 0;
-                        xml_node_struct* cursor = xmldoc;
+                        xml_node_struct* cursor = root;
                         char_t* mark = s;
 
@@ -2403 +3176 @@
                                                 cursor->name = s;
 
-                                                PUGI__SCANWHILE(PUGI__IS_CHARTYPE(*s, ct_symbol)); // Scan for a terminator.
+                                                PUGI__SCANWHILE_UNROLL(PUGI__IS_CHARTYPE(ss, ct_symbol)); // Scan for a terminator.
                                                 PUGI__ENDSEG(); // Save char in 'ch', terminate & step over.
 
@@ -2419 +3192 @@
                                                                 if (PUGI__IS_CHARTYPE(*s, ct_start_symbol)) // <... #...
                                                                 {
-                                                                        xml_attribute_struct* a = append_attribute_ll(cursor, alloc); // Make space for this attribute.
+                                                                        xml_attribute_struct* a = append_new_attribute(cursor, alloc); // Make space for this attribute.
                                                                         if (!a) PUGI__THROW_ERROR(status_out_of_memory, s);
 
                                                                         a->name = s; // Save the offset.
 
-                                                                        PUGI__SCANWHILE(PUGI__IS_CHARTYPE(*s, ct_symbol)); // Scan for a terminator.
-                                                                        PUGI__CHECK_ERROR(status_bad_attribute, s); //$ redundant, left for performance
-
+                                                                        PUGI__SCANWHILE_UNROLL(PUGI__IS_CHARTYPE(ss, ct_symbol)); // Scan for a terminator.
                                                                         PUGI__ENDSEG(); // Save char in 'ch', terminate & step over.
-                                                                        PUGI__CHECK_ERROR(status_bad_attribute, s); //$ redundant, left for performance
 
                                                                         if (PUGI__IS_CHARTYPE(ch, ct_space))
                                                                         {
                                                                                 PUGI__SKIPWS(); // Eat any whitespace.
-                                                                                PUGI__CHECK_ERROR(status_bad_attribute, s); //$ redundant, left for performance
 
                                                                                 ch = *s;
@@ -2496 +3265 @@
                                                 else if (ch == '/') // '<#.../'
                                                 {
-                                                        if (!ENDSWITH(*s, '>')) PUGI__THROW_ERROR(status_bad_start_element, s);
+                                                        if (!PUGI__ENDSWITH(*s, '>')) PUGI__THROW_ERROR(status_bad_start_element, s);
 
                                                         PUGI__POPNODE(); // Pop.
@@ -2549 +3318 @@
 
                                                 assert(cursor);
-                                                if ((cursor->header & xml_memory_page_type_mask) + 1 == node_declaration) goto LOC_ATTRIBUTES;
+                                                if (PUGI__NODETYPE(cursor) == node_declaration) goto LOC_ATTRIBUTES;
                                         }
                                         else if (*s == '!') // '<!...'
@@ -2565 +3334 @@
                                         PUGI__SKIPWS(); // Eat whitespace if no genuine PCDATA here.
 
-                                        if (*s == '<')
+                                        if (*s == '<' || !*s)
                                         {
                                                 // We skipped some whitespace characters because otherwise we would take the tag branch instead of PCDATA one
                                                 assert(mark != s);
 
-                                                if (!PUGI__OPTSET(parse_ws_pcdata | parse_ws_pcdata_single))
+                                                if (!PUGI__OPTSET(parse_ws_pcdata | parse_ws_pcdata_single) || PUGI__OPTSET(parse_trim_pcdata))
                                                 {
                                                         continue;
@@ -2576 +3345 @@
                                                 else if (PUGI__OPTSET(parse_ws_pcdata_single))
                                                 {
-                                                        if (s[1] != '/' || cursor->first_child) continue;
+                                                        if (s[0] != '<' || s[1] != '/' || cursor->first_child) continue;
                                                 }
                                         }
 
-                                        s = mark;
+                                        if (!PUGI__OPTSET(parse_trim_pcdata))
+                                                s = mark;
                                                         
-                                        if (cursor->parent)
+                                        if (cursor->parent || PUGI__OPTSET(parse_fragment))
                                         {
                                                 PUGI__PUSHNODE(node_pcdata); // Append a new node on the tree.
@@ -2607 +3377 @@
 
                         // check that last tag is closed
-                        if (cursor != xmldoc) PUGI__THROW_ERROR(status_end_element_mismatch, s);
+                        if (cursor != root) PUGI__THROW_ERROR(status_end_element_mismatch, s);
 
                         return s;
                 }
 
-                static xml_parse_result parse(char_t* buffer, size_t length, xml_node_struct* root, unsigned int optmsk)
-                {
-                        xml_document_struct* xmldoc = static_cast<xml_document_struct*>(root);
-
-                        // store buffer for offset_debug
-                        xmldoc->buffer = buffer;
-
+        #ifdef PUGIXML_WCHAR_MODE
+                static char_t* parse_skip_bom(char_t* s)
+                {
+                        unsigned int bom = 0xfeff;
+                        return (s[0] == static_cast<wchar_t>(bom)) ? s + 1 : s;
+                }
+        #else
+                static char_t* parse_skip_bom(char_t* s)
+                {
+                        return (s[0] == '\xef' && s[1] == '\xbb' && s[2] == '\xbf') ? s + 3 : s;
+                }
+        #endif
+
+                static bool has_element_node_siblings(xml_node_struct* node)
+                {
+                        while (node)
+                        {
+                                if (PUGI__NODETYPE(node) == node_element) return true;
+
+                                node = node->next_sibling;
+                        }
+
+                        return false;
+                }
+
+                static xml_parse_result parse(char_t* buffer, size_t length, xml_document_struct* xmldoc, xml_node_struct* root, unsigned int optmsk)
+                {
                         // early-out for empty documents
-                        if (length == 0) return make_parse_result(status_ok);
-
+                        if (length == 0)
+                                return make_parse_result(PUGI__OPTSET(parse_fragment) ? status_ok : status_no_document_element);
+
+                        // get last child of the root before parsing
+                        xml_node_struct* last_root_child = root->first_child ? root->first_child->prev_sibling_c + 0 : 0;
+        
                         // create parser on stack
-                        xml_parser parser(*xmldoc);
+                        xml_parser parser(static_cast<xml_allocator*>(xmldoc));
 
                         // save last character and make buffer zero-terminated (speeds up parsing)
@@ -2629 +3423 @@
                         buffer[length - 1] = 0;
                         
+                        // skip BOM to make sure it does not end up as part of parse output
+                        char_t* buffer_data = parse_skip_bom(buffer);
+
                         // perform actual parsing
-                        parser.parse(buffer, xmldoc, optmsk, endch);
+                        parser.parse_tree(buffer_data, root, optmsk, endch);
 
                         xml_parse_result result = make_parse_result(parser.error_status, parser.error_offset ? parser.error_offset - buffer : 0);
                         assert(result.offset >= 0 && static_cast<size_t>(result.offset) <= length);
 
-                        // update allocator state
-                        *static_cast<xml_allocator*>(xmldoc) = parser.alloc;
-
-                        // since we removed last character, we have to handle the only possible false positive
-                        if (result && endch == '<')
-                        {
-                                // there's no possible well-formed document with < at the end
-                                return make_parse_result(status_unrecognized_tag, length);
+                        if (result)
+                        {
+                                // since we removed last character, we have to handle the only possible false positive (stray <)
+                                if (endch == '<')
+                                        return make_parse_result(status_unrecognized_tag, length - 1);
+
+                                // check if there are any element nodes parsed
+                                xml_node_struct* first_root_child_parsed = last_root_child ? last_root_child->next_sibling + 0 : root->first_child+ 0;
+
+                                if (!PUGI__OPTSET(parse_fragment) && !has_element_node_siblings(first_root_child_parsed))
+                                        return make_parse_result(status_no_document_element, length - 1);
+                        }
+                        else
+                        {
+                                // roll back offset if it occurs on a null terminator in the source buffer
+                                if (result.offset > 0 && static_cast<size_t>(result.offset) == length - 1 && endch == 0)
+                                        result.offset--;
                         }
 
@@ -2677 +3483 @@
         }
 
+        template <typename D, typename T> PUGI__FN size_t convert_buffer_output_generic(typename T::value_type dest, const char_t* data, size_t length, D, T)
+        {
+                PUGI__STATIC_ASSERT(sizeof(char_t) == sizeof(typename D::type));
+
+                typename T::value_type end = D::process(reinterpret_cast<const typename D::type*>(data), length, dest, T());
+
+                return static_cast<size_t>(end - dest) * sizeof(*dest);
+        }
+
+        template <typename D, typename T> PUGI__FN size_t convert_buffer_output_generic(typename T::value_type dest, const char_t* data, size_t length, D, T, bool opt_swap)
+        {
+                PUGI__STATIC_ASSERT(sizeof(char_t) == sizeof(typename D::type));
+
+                typename T::value_type end = D::process(reinterpret_cast<const typename D::type*>(data), length, dest, T());
+
+                if (opt_swap)
+                {
+                        for (typename T::value_type i = dest; i != end; ++i)
+                                *i = endian_swap(*i);
+                }
+
+                return static_cast<size_t>(end - dest) * sizeof(*dest);
+        }
+
 #ifdef PUGIXML_WCHAR_MODE
         PUGI__FN size_t get_valid_length(const char_t* data, size_t length)
         {
-                assert(length > 0);
+                if (length < 1) return 0;
 
                 // discard last character if it's the lead of a surrogate pair 
@@ -2686 +3516 @@
         }
 
-        PUGI__FN size_t convert_buffer(char_t* r_char, uint8_t* r_u8, uint16_t* r_u16, uint32_t* r_u32, const char_t* data, size_t length, xml_encoding encoding)
+        PUGI__FN size_t convert_buffer_output(char_t* r_char, uint8_t* r_u8, uint16_t* r_u16, uint32_t* r_u32, const char_t* data, size_t length, xml_encoding encoding)
         {
                 // only endian-swapping is required
@@ -2698 +3528 @@
                 // convert to utf8
                 if (encoding == encoding_utf8)
-                {
-                        uint8_t* dest = r_u8;
-                        uint8_t* end = utf_decoder<utf8_writer>::decode_wchar_block(data, length, dest);
-
-                        return static_cast<size_t>(end - dest);
-                }
+                        return convert_buffer_output_generic(r_u8, data, length, wchar_decoder(), utf8_writer());
 
                 // convert to utf16
                 if (encoding == encoding_utf16_be || encoding == encoding_utf16_le)
                 {
-                        uint16_t* dest = r_u16;
-
-                        // convert to native utf16
-                        uint16_t* end = utf_decoder<utf16_writer>::decode_wchar_block(data, length, dest);
-
-                        // swap if necessary
                         xml_encoding native_encoding = is_little_endian() ? encoding_utf16_le : encoding_utf16_be;
 
-                        if (native_encoding != encoding) convert_utf_endian_swap(dest, dest, static_cast<size_t>(end - dest));
-
-                        return static_cast<size_t>(end - dest) * sizeof(uint16_t);
+                        return convert_buffer_output_generic(r_u16, data, length, wchar_decoder(), utf16_writer(), native_encoding != encoding);
                 }
 
@@ -2724 +3541 @@
                 if (encoding == encoding_utf32_be || encoding == encoding_utf32_le)
                 {
-                        uint32_t* dest = r_u32;
-
-                        // convert to native utf32
-                        uint32_t* end = utf_decoder<utf32_writer>::decode_wchar_block(data, length, dest);
-
-                        // swap if necessary
                         xml_encoding native_encoding = is_little_endian() ? encoding_utf32_le : encoding_utf32_be;
 
-                        if (native_encoding != encoding) convert_utf_endian_swap(dest, dest, static_cast<size_t>(end - dest));
-
-                        return static_cast<size_t>(end - dest) * sizeof(uint32_t);
+                        return convert_buffer_output_generic(r_u32, data, length, wchar_decoder(), utf32_writer(), native_encoding != encoding);
                 }
 
                 // convert to latin1
                 if (encoding == encoding_latin1)
-                {
-                        uint8_t* dest = r_u8;
-                        uint8_t* end = utf_decoder<latin1_writer>::decode_wchar_block(data, length, dest);
-
-                        return static_cast<size_t>(end - dest);
-                }
+                        return convert_buffer_output_generic(r_u8, data, length, wchar_decoder(), latin1_writer());
 
                 assert(!"Invalid encoding");
@@ -2752 +3556 @@
         PUGI__FN size_t get_valid_length(const char_t* data, size_t length)
         {
-                assert(length > 4);
+                if (length < 5) return 0;
 
                 for (size_t i = 1; i <= 4; ++i)
@@ -2766 +3570 @@
         }
 
-        PUGI__FN size_t convert_buffer(char_t* /* r_char */, uint8_t* r_u8, uint16_t* r_u16, uint32_t* r_u32, const char_t* data, size_t length, xml_encoding encoding)
+        PUGI__FN size_t convert_buffer_output(char_t* /* r_char */, uint8_t* r_u8, uint16_t* r_u16, uint32_t* r_u32, const char_t* data, size_t length, xml_encoding encoding)
         {
                 if (encoding == encoding_utf16_be || encoding == encoding_utf16_le)
                 {
-                        uint16_t* dest = r_u16;
-
-                        // convert to native utf16
-                        uint16_t* end = utf_decoder<utf16_writer>::decode_utf8_block(reinterpret_cast<const uint8_t*>(data), length, dest);
-
-                        // swap if necessary
                         xml_encoding native_encoding = is_little_endian() ? encoding_utf16_le : encoding_utf16_be;
 
-                        if (native_encoding != encoding) convert_utf_endian_swap(dest, dest, static_cast<size_t>(end - dest));
-
-                        return static_cast<size_t>(end - dest) * sizeof(uint16_t);
+                        return convert_buffer_output_generic(r_u16, data, length, utf8_decoder(), utf16_writer(), native_encoding != encoding);
                 }
 
                 if (encoding == encoding_utf32_be || encoding == encoding_utf32_le)
                 {
-                        uint32_t* dest = r_u32;
-
-                        // convert to native utf32
-                        uint32_t* end = utf_decoder<utf32_writer>::decode_utf8_block(reinterpret_cast<const uint8_t*>(data), length, dest);
-
-                        // swap if necessary
                         xml_encoding native_encoding = is_little_endian() ? encoding_utf32_le : encoding_utf32_be;
 
-                        if (native_encoding != encoding) convert_utf_endian_swap(dest, dest, static_cast<size_t>(end - dest));
-
-                        return static_cast<size_t>(end - dest) * sizeof(uint32_t);
+                        return convert_buffer_output_generic(r_u32, data, length, utf8_decoder(), utf32_writer(), native_encoding != encoding);
                 }
 
                 if (encoding == encoding_latin1)
-                {
-                        uint8_t* dest = r_u8;
-                        uint8_t* end = utf_decoder<latin1_writer>::decode_utf8_block(reinterpret_cast<const uint8_t*>(data), length, dest);
-
-                        return static_cast<size_t>(end - dest);
-                }
+                        return convert_buffer_output_generic(r_u8, data, length, utf8_decoder(), latin1_writer());
 
                 assert(!"Invalid encoding");
@@ -2822 +3605 @@
                 }
 
-                ~xml_buffered_writer()
-                {
-                        flush();
-                }
-
-                void flush()
+                size_t flush()
                 {
                         flush(buffer, bufsize);
                         bufsize = 0;
+                        return 0;
                 }
 
@@ -2843 +3622 @@
                         {
                                 // convert chunk
-                                size_t result = convert_buffer(scratch.data_char, scratch.data_u8, scratch.data_u16, scratch.data_u32, data, size, encoding);
+                                size_t result = convert_buffer_output(scratch.data_char, scratch.data_u8, scratch.data_u16, scratch.data_u32, data, size, encoding);
                                 assert(result <= sizeof(scratch));
 
@@ -2851 +3630 @@
                 }
 
-                void write(const char_t* data, size_t length)
-                {
-                        if (bufsize + length > bufcapacity)
-                        {
-                                // flush the remaining buffer contents
-                                flush();
-
-                                // handle large chunks
-                                if (length > bufcapacity)
-                                {
-                                        if (encoding == get_write_native_encoding())
-                                        {
-                                                // fast path, can just write data chunk
-                                                writer.write(data, length * sizeof(char_t));
-                                                return;
-                                        }
-
-                                        // need to convert in suitable chunks
-                                        while (length > bufcapacity)
-                                        {
-                                                // get chunk size by selecting such number of characters that are guaranteed to fit into scratch buffer
-                                                // and form a complete codepoint sequence (i.e. discard start of last codepoint if necessary)
-                                                size_t chunk_size = get_valid_length(data, bufcapacity);
-
-                                                // convert chunk and write
-                                                flush(data, chunk_size);
-
-                                                // iterate
-                                                data += chunk_size;
-                                                length -= chunk_size;
-                                        }
-
-                                        // small tail is copied below
-                                        bufsize = 0;
-                                }
+                void write_direct(const char_t* data, size_t length)
+                {
+                        // flush the remaining buffer contents
+                        flush();
+
+                        // handle large chunks
+                        if (length > bufcapacity)
+                        {
+                                if (encoding == get_write_native_encoding())
+                                {
+                                        // fast path, can just write data chunk
+                                        writer.write(data, length * sizeof(char_t));
+                                        return;
+                                }
+
+                                // need to convert in suitable chunks
+                                while (length > bufcapacity)
+                                {
+                                        // get chunk size by selecting such number of characters that are guaranteed to fit into scratch buffer
+                                        // and form a complete codepoint sequence (i.e. discard start of last codepoint if necessary)
+                                        size_t chunk_size = get_valid_length(data, bufcapacity);
+                                        assert(chunk_size);
+
+                                        // convert chunk and write
+                                        flush(data, chunk_size);
+
+                                        // iterate
+                                        data += chunk_size;
+                                        length -= chunk_size;
+                                }
+
+                                // small tail is copied below
+                                bufsize = 0;
                         }
 
@@ -2892 +3669 @@
                 }
 
-                void write(const char_t* data)
-                {
-                        write(data, strlength(data));
+                void write_buffer(const char_t* data, size_t length)
+                {
+                        size_t offset = bufsize;
+
+                        if (offset + length <= bufcapacity)
+                        {
+                                memcpy(buffer + offset, data, length * sizeof(char_t));
+                                bufsize = offset + length;
+                        }
+                        else
+                        {
+                                write_direct(data, length);
+                        }
+                }
+
+                void write_string(const char_t* data)
+                {
+                        // write the part of the string that fits in the buffer
+                        size_t offset = bufsize;
+
+                        while (*data && offset < bufcapacity)
+                                buffer[offset++] = *data++;
+
+                        // write the rest
+                        if (offset < bufcapacity)
+                        {
+                                bufsize = offset;
+                        }
+                        else
+                        {
+                                // backtrack a bit if we have split the codepoint
+                                size_t length = offset - bufsize;
+                                size_t extra = length - get_valid_length(data - length, length);
+
+                                bufsize = offset - extra;
+
+                                write_direct(data - extra, strlength(data) + extra);
+                        }
                 }
 
                 void write(char_t d0)
                 {
-                        if (bufsize + 1 > bufcapacity) flush();
-
-                        buffer[bufsize + 0] = d0;
-                        bufsize += 1;
+                        size_t offset = bufsize;
+                        if (offset > bufcapacity - 1) offset = flush();
+
+                        buffer[offset + 0] = d0;
+                        bufsize = offset + 1;
                 }
 
                 void write(char_t d0, char_t d1)
                 {
-                        if (bufsize + 2 > bufcapacity) flush();
-
-                        buffer[bufsize + 0] = d0;
-                        buffer[bufsize + 1] = d1;
-                        bufsize += 2;
+                        size_t offset = bufsize;
+                        if (offset > bufcapacity - 2) offset = flush();
+
+                        buffer[offset + 0] = d0;
+                        buffer[offset + 1] = d1;
+                        bufsize = offset + 2;
                 }
 
                 void write(char_t d0, char_t d1, char_t d2)
                 {
-                        if (bufsize + 3 > bufcapacity) flush();
-
-                        buffer[bufsize + 0] = d0;
-                        buffer[bufsize + 1] = d1;
-                        buffer[bufsize + 2] = d2;
-                        bufsize += 3;
+                        size_t offset = bufsize;
+                        if (offset > bufcapacity - 3) offset = flush();
+
+                        buffer[offset + 0] = d0;
+                        buffer[offset + 1] = d1;
+                        buffer[offset + 2] = d2;
+                        bufsize = offset + 3;
                 }
 
                 void write(char_t d0, char_t d1, char_t d2, char_t d3)
                 {
-                        if (bufsize + 4 > bufcapacity) flush();
-
-                        buffer[bufsize + 0] = d0;
-                        buffer[bufsize + 1] = d1;
-                        buffer[bufsize + 2] = d2;
-                        buffer[bufsize + 3] = d3;
-                        bufsize += 4;
+                        size_t offset = bufsize;
+                        if (offset > bufcapacity - 4) offset = flush();
+
+                        buffer[offset + 0] = d0;
+                        buffer[offset + 1] = d1;
+                        buffer[offset + 2] = d2;
+                        buffer[offset + 3] = d3;
+                        bufsize = offset + 4;
                 }
 
                 void write(char_t d0, char_t d1, char_t d2, char_t d3, char_t d4)
                 {
-                        if (bufsize + 5 > bufcapacity) flush();
-
-                        buffer[bufsize + 0] = d0;
-                        buffer[bufsize + 1] = d1;
-                        buffer[bufsize + 2] = d2;
-                        buffer[bufsize + 3] = d3;
-                        buffer[bufsize + 4] = d4;
-                        bufsize += 5;
+                        size_t offset = bufsize;
+                        if (offset > bufcapacity - 5) offset = flush();
+
+                        buffer[offset + 0] = d0;
+                        buffer[offset + 1] = d1;
+                        buffer[offset + 2] = d2;
+                        buffer[offset + 3] = d3;
+                        buffer[offset + 4] = d4;
+                        bufsize = offset + 5;
                 }
 
                 void write(char_t d0, char_t d1, char_t d2, char_t d3, char_t d4, char_t d5)
                 {
-                        if (bufsize + 6 > bufcapacity) flush();
-
-                        buffer[bufsize + 0] = d0;
-                        buffer[bufsize + 1] = d1;
-                        buffer[bufsize + 2] = d2;
-                        buffer[bufsize + 3] = d3;
-                        buffer[bufsize + 4] = d4;
-                        buffer[bufsize + 5] = d5;
-                        bufsize += 6;
+                        size_t offset = bufsize;
+                        if (offset > bufcapacity - 6) offset = flush();
+
+                        buffer[offset + 0] = d0;
+                        buffer[offset + 1] = d1;
+                        buffer[offset + 2] = d2;
+                        buffer[offset + 3] = d3;
+                        buffer[offset + 4] = d4;
+                        buffer[offset + 5] = d5;
+                        bufsize = offset + 6;
                 }
 
@@ -2997 +3815 @@
                         
                         // While *s is a usual symbol
-                        while (!PUGI__IS_CHARTYPEX(*s, type)) ++s;
+                        PUGI__SCANWHILE_UNROLL(!PUGI__IS_CHARTYPEX(ss, type));
                 
-                        writer.write(prev, static_cast<size_t>(s - prev));
+                        writer.write_buffer(prev, static_cast<size_t>(s - prev));
 
                         switch (*s)
@@ -3034 +3852 @@
         {
                 if (flags & format_no_escapes)
-                        writer.write(s);
+                        writer.write_string(s);
                 else
                         text_output_escaped(writer, s, type);
@@ -3054 +3872 @@
                         if (*s) s += 2;
 
-                        writer.write(prev, static_cast<size_t>(s - prev));
+                        writer.write_buffer(prev, static_cast<size_t>(s - prev));
 
                         writer.write(']', ']', '>');
@@ -3061 +3879 @@
         }
 
-        PUGI__FN void node_output_attributes(xml_buffered_writer& writer, const xml_node& node, unsigned int flags)
+        PUGI__FN void text_output_indent(xml_buffered_writer& writer, const char_t* indent, size_t indent_length, unsigned int depth)
+        {
+                switch (indent_length)
+                {
+                case 1:
+                {
+                        for (unsigned int i = 0; i < depth; ++i)
+                                writer.write(indent[0]);
+                        break;
+                }
+
+                case 2:
+                {
+                        for (unsigned int i = 0; i < depth; ++i)
+                                writer.write(indent[0], indent[1]);
+                        break;
+                }
+
+                case 3:
+                {
+                        for (unsigned int i = 0; i < depth; ++i)
+                                writer.write(indent[0], indent[1], indent[2]);
+                        break;
+                }
+
+                case 4:
+                {
+                        for (unsigned int i = 0; i < depth; ++i)
+                                writer.write(indent[0], indent[1], indent[2], indent[3]);
+                        break;
+                }
+
+                default:
+                {
+                        for (unsigned int i = 0; i < depth; ++i)
+                                writer.write_buffer(indent, indent_length);
+                }
+                }
+        }
+
+        PUGI__FN void node_output_comment(xml_buffered_writer& writer, const char_t* s)
+        {
+                writer.write('<', '!', '-', '-');
+
+                while (*s)
+                {
+                        const char_t* prev = s;
+
+                        // look for -\0 or -- sequence - we can't output it since -- is illegal in comment body
+                        while (*s && !(s[0] == '-' && (s[1] == '-' || s[1] == 0))) ++s;
+
+                        writer.write_buffer(prev, static_cast<size_t>(s - prev));
+
+                        if (*s)
+                        {
+                                assert(*s == '-');
+
+                                writer.write('-', ' ');
+                                ++s;
+                        }
+                }
+
+                writer.write('-', '-', '>');
+        }
+
+        PUGI__FN void node_output_pi_value(xml_buffered_writer& writer, const char_t* s)
+        {
+                while (*s)
+                {
+                        const char_t* prev = s;
+
+                        // look for ?> sequence - we can't output it since ?> terminates PI
+                        while (*s && !(s[0] == '?' && s[1] == '>')) ++s;
+
+                        writer.write_buffer(prev, static_cast<size_t>(s - prev));
+
+                        if (*s)
+                        {
+                                assert(s[0] == '?' && s[1] == '>');
+
+                                writer.write('?', ' ', '>');
+                                s += 2;
+                        }
+                }
+        }
+
+        PUGI__FN void node_output_attributes(xml_buffered_writer& writer, xml_node_struct* node, const char_t* indent, size_t indent_length, unsigned int flags, unsigned int depth)
         {
                 const char_t* default_name = PUGIXML_TEXT(":anonymous");
 
-                for (xml_attribute a = node.first_attribute(); a; a = a.next_attribute())
-                {
-                        writer.write(' ');
-                        writer.write(a.name()[0] ? a.name() : default_name);
+                for (xml_attribute_struct* a = node->first_attribute; a; a = a->next_attribute)
+                {
+                        if ((flags & (format_indent_attributes | format_raw)) == format_indent_attributes)
+                        {
+                                writer.write('\n');
+
+                                text_output_indent(writer, indent, indent_length, depth + 1);
+                        }
+                        else
+                        {
+                                writer.write(' ');
+                        }
+
+                        writer.write_string(a->name ? a->name + 0 : default_name);
                         writer.write('=', '"');
 
-                        text_output(writer, a.value(), ctx_special_attr, flags);
+                        if (a->value)
+                                text_output(writer, a->value, ctx_special_attr, flags);
 
                         writer.write('"');
@@ -3077 +3992 @@
         }
 
-        PUGI__FN void node_output(xml_buffered_writer& writer, const xml_node& node, const char_t* indent, unsigned int flags, unsigned int depth)
+        PUGI__FN bool node_output_start(xml_buffered_writer& writer, xml_node_struct* node, const char_t* indent, size_t indent_length, unsigned int flags, unsigned int depth)
         {
                 const char_t* default_name = PUGIXML_TEXT(":anonymous");
-
-                if ((flags & format_indent) != 0 && (flags & format_raw) == 0)
-                        for (unsigned int i = 0; i < depth; ++i) writer.write(indent);
-
-                switch (node.type())
-                {
-                case node_document:
-                {
-                        for (xml_node n = node.first_child(); n; n = n.next_sibling())
-                                node_output(writer, n, indent, flags, depth);
-                        break;
-                }
-                        
-                case node_element:
-                {
-                        const char_t* name = node.name()[0] ? node.name() : default_name;
-
-                        writer.write('<');
-                        writer.write(name);
-
-                        node_output_attributes(writer, node, flags);
-
-                        if (flags & format_raw)
-                        {
-                                if (!node.first_child())
-                                        writer.write(' ', '/', '>');
+                const char_t* name = node->name ? node->name + 0 : default_name;
+
+                writer.write('<');
+                writer.write_string(name);
+
+                if (node->first_attribute)
+                        node_output_attributes(writer, node, indent, indent_length, flags, depth);
+
+                if (!node->first_child)
+                {
+                        writer.write(' ', '/', '>');
+
+                        return false;
+                }
+                else
+                {
+                        writer.write('>');
+
+                        return true;
+                }
+        }
+
+        PUGI__FN void node_output_end(xml_buffered_writer& writer, xml_node_struct* node)
+        {
+                const char_t* default_name = PUGIXML_TEXT(":anonymous");
+                const char_t* name = node->name ? node->name + 0 : default_name;
+
+                writer.write('<', '/');
+                writer.write_string(name);
+                writer.write('>');
+        }
+
+        PUGI__FN void node_output_simple(xml_buffered_writer& writer, xml_node_struct* node, unsigned int flags)
+        {
+                const char_t* default_name = PUGIXML_TEXT(":anonymous");
+
+                switch (PUGI__NODETYPE(node))
+                {
+                        case node_pcdata:
+                                text_output(writer, node->value ? node->value + 0 : PUGIXML_TEXT(""), ctx_special_pcdata, flags);
+                                break;
+
+                        case node_cdata:
+                                text_output_cdata(writer, node->value ? node->value + 0 : PUGIXML_TEXT(""));
+                                break;
+
+                        case node_comment:
+                                node_output_comment(writer, node->value ? node->value + 0 : PUGIXML_TEXT(""));
+                                break;
+
+                        case node_pi:
+                                writer.write('<', '?');
+                                writer.write_string(node->name ? node->name + 0 : default_name);
+
+                                if (node->value)
+                                {
+                                        writer.write(' ');
+                                        node_output_pi_value(writer, node->value);
+                                }
+
+                                writer.write('?', '>');
+                                break;
+
+                        case node_declaration:
+                                writer.write('<', '?');
+                                writer.write_string(node->name ? node->name + 0 : default_name);
+                                node_output_attributes(writer, node, PUGIXML_TEXT(""), 0, flags | format_raw, 0);
+                                writer.write('?', '>');
+                                break;
+
+                        case node_doctype:
+                                writer.write('<', '!', 'D', 'O', 'C');
+                                writer.write('T', 'Y', 'P', 'E');
+
+                                if (node->value)
+                                {
+                                        writer.write(' ');
+                                        writer.write_string(node->value);
+                                }
+
+                                writer.write('>');
+                                break;
+
+                        default:
+                                assert(!"Invalid node type");
+                }
+        }
+
+        enum indent_flags_t
+        {
+                indent_newline = 1,
+                indent_indent = 2
+        };
+
+        PUGI__FN void node_output(xml_buffered_writer& writer, xml_node_struct* root, const char_t* indent, unsigned int flags, unsigned int depth)
+        {
+                size_t indent_length = ((flags & (format_indent | format_indent_attributes)) && (flags & format_raw) == 0) ? strlength(indent) : 0;
+                unsigned int indent_flags = indent_indent;
+
+                xml_node_struct* node = root;
+
+                do
+                {
+                        assert(node);
+
+                        // begin writing current node
+                        if (PUGI__NODETYPE(node) == node_pcdata || PUGI__NODETYPE(node) == node_cdata)
+                        {
+                                node_output_simple(writer, node, flags);
+
+                                indent_flags = 0;
+                        }
+                        else
+                        {
+                                if ((indent_flags & indent_newline) && (flags & format_raw) == 0)
+                                        writer.write('\n');
+
+                                if ((indent_flags & indent_indent) && indent_length)
+                                        text_output_indent(writer, indent, indent_length, depth);
+
+                                if (PUGI__NODETYPE(node) == node_element)
+                                {
+                                        indent_flags = indent_newline | indent_indent;
+
+                                        if (node_output_start(writer, node, indent, indent_length, flags, depth))
+                                        {
+                                                node = node->first_child;
+                                                depth++;
+                                                continue;
+                                        }
+                                }
+                                else if (PUGI__NODETYPE(node) == node_document)
+                                {
+                                        indent_flags = indent_indent;
+
+                                        if (node->first_child)
+                                        {
+                                                node = node->first_child;
+                                                continue;
+                                        }
+                                }
                                 else
                                 {
-                                        writer.write('>');
-
-                                        for (xml_node n = node.first_child(); n; n = n.next_sibling())
-                                                node_output(writer, n, indent, flags, depth + 1);
-
-                                        writer.write('<', '/');
-                                        writer.write(name);
-                                        writer.write('>');
-                                }
-                        }
-                        else if (!node.first_child())
-                                writer.write(' ', '/', '>', '\n');
-                        else if (node.first_child() == node.last_child() && (node.first_child().type() == node_pcdata || node.first_child().type() == node_cdata))
-                        {
-                                writer.write('>');
-
-                                if (node.first_child().type() == node_pcdata)
-                                        text_output(writer, node.first_child().value(), ctx_special_pcdata, flags);
-                                else
-                                        text_output_cdata(writer, node.first_child().value());
-
-                                writer.write('<', '/');
-                                writer.write(name);
-                                writer.write('>', '\n');
-                        }
-                        else
-                        {
-                                writer.write('>', '\n');
-                                
-                                for (xml_node n = node.first_child(); n; n = n.next_sibling())
-                                        node_output(writer, n, indent, flags, depth + 1);
-
-                                if ((flags & format_indent) != 0 && (flags & format_raw) == 0)
-                                        for (unsigned int i = 0; i < depth; ++i) writer.write(indent);
-                                
-                                writer.write('<', '/');
-                                writer.write(name);
-                                writer.write('>', '\n');
-                        }
-
-                        break;
-                }
-                
-                case node_pcdata:
-                        text_output(writer, node.value(), ctx_special_pcdata, flags);
-                        if ((flags & format_raw) == 0) writer.write('\n');
-                        break;
-
-                case node_cdata:
-                        text_output_cdata(writer, node.value());
-                        if ((flags & format_raw) == 0) writer.write('\n');
-                        break;
-
-                case node_comment:
-                        writer.write('<', '!', '-', '-');
-                        writer.write(node.value());
-                        writer.write('-', '-', '>');
-                        if ((flags & format_raw) == 0) writer.write('\n');
-                        break;
-
-                case node_pi:
-                case node_declaration:
-                        writer.write('<', '?');
-                        writer.write(node.name()[0] ? node.name() : default_name);
-
-                        if (node.type() == node_declaration)
-                        {
-                                node_output_attributes(writer, node, flags);
-                        }
-                        else if (node.value()[0])
-                        {
-                                writer.write(' ');
-                                writer.write(node.value());
-                        }
-
-                        writer.write('?', '>');
-                        if ((flags & format_raw) == 0) writer.write('\n');
-                        break;
-
-                case node_doctype:
-                        writer.write('<', '!', 'D', 'O', 'C');
-                        writer.write('T', 'Y', 'P', 'E');
-
-                        if (node.value()[0])
-                        {
-                                writer.write(' ');
-                                writer.write(node.value());
-                        }
-
-                        writer.write('>');
-                        if ((flags & format_raw) == 0) writer.write('\n');
-                        break;
-
-                default:
-                        assert(!"Invalid node type");
-                }
-        }
-
-        inline bool has_declaration(const xml_node& node)
-        {
-                for (xml_node child = node.first_child(); child; child = child.next_sibling())
-                {
-                        xml_node_type type = child.type();
+                                        node_output_simple(writer, node, flags);
+
+                                        indent_flags = indent_newline | indent_indent;
+                                }
+                        }
+
+                        // continue to the next node
+                        while (node != root)
+                        {
+                                if (node->next_sibling)
+                                {
+                                        node = node->next_sibling;
+                                        break;
+                                }
+
+                                node = node->parent;
+
+                                // write closing node
+                                if (PUGI__NODETYPE(node) == node_element)
+                                {
+                                        depth--;
+
+                                        if ((indent_flags & indent_newline) && (flags & format_raw) == 0)
+                                                writer.write('\n');
+
+                                        if ((indent_flags & indent_indent) && indent_length)
+                                                text_output_indent(writer, indent, indent_length, depth);
+
+                                        node_output_end(writer, node);
+
+                                        indent_flags = indent_newline | indent_indent;
+                                }
+                        }
+                }
+                while (node != root);
+
+                if ((indent_flags & indent_newline) && (flags & format_raw) == 0)
+                        writer.write('\n');
+        }
+
+        PUGI__FN bool has_declaration(xml_node_struct* node)
+        {
+                for (xml_node_struct* child = node->first_child; child; child = child->next_sibling)
+                {
+                        xml_node_type type = PUGI__NODETYPE(child);
 
                         if (type == node_declaration) return true;
@@ -3219 +4191 @@
         }
 
-        inline bool allow_insert_child(xml_node_type parent, xml_node_type child)
+        PUGI__FN bool is_attribute_of(xml_attribute_struct* attr, xml_node_struct* node)
+        {
+                for (xml_attribute_struct* a = node->first_attribute; a; a = a->next_attribute)
+                        if (a == attr)
+                                return true;
+
+                return false;
+        }
+
+        PUGI__FN bool allow_insert_attribute(xml_node_type parent)
+        {
+                return parent == node_element || parent == node_declaration;
+        }
+
+        PUGI__FN bool allow_insert_child(xml_node_type parent, xml_node_type child)
         {
                 if (parent != node_document && parent != node_element) return false;
@@ -3228 +4214 @@
         }
 
-        PUGI__FN void recursive_copy_skip(xml_node& dest, const xml_node& source, const xml_node& skip)
-        {
-                assert(dest.type() == source.type());
-
-                switch (source.type())
-                {
-                case node_element:
-                {
-                        dest.set_name(source.name());
-
-                        for (xml_attribute a = source.first_attribute(); a; a = a.next_attribute())
-                                dest.append_attribute(a.name()).set_value(a.value());
-
-                        for (xml_node c = source.first_child(); c; c = c.next_sibling())
-                        {
-                                if (c == skip) continue;
-
-                                xml_node cc = dest.append_child(c.type());
-                                assert(cc);
-
-                                recursive_copy_skip(cc, c, skip);
-                        }
-
-                        break;
-                }
-
-                case node_pcdata:
-                case node_cdata:
-                case node_comment:
-                case node_doctype:
-                        dest.set_value(source.value());
-                        break;
-
-                case node_pi:
-                        dest.set_name(source.name());
-                        dest.set_value(source.value());
-                        break;
-
-                case node_declaration:
-                {
-                        dest.set_name(source.name());
-
-                        for (xml_attribute a = source.first_attribute(); a; a = a.next_attribute())
-                                dest.append_attribute(a.name()).set_value(a.value());
-
-                        break;
-                }
-
-                default:
-                        assert(!"Invalid node type");
-                }
+        PUGI__FN bool allow_move(xml_node parent, xml_node child)
+        {
+                // check that child can be a child of parent
+                if (!allow_insert_child(parent.type(), child.type()))
+                        return false;
+
+                // check that node is not moved between documents
+                if (parent.root() != child.root())
+                        return false;
+
+                // check that new parent is not in the child subtree
+                xml_node cur = parent;
+
+                while (cur)
+                {
+                        if (cur == child)
+                                return false;
+
+                        cur = cur.parent();
+                }
+
+                return true;
+        }
+
+        template <typename String, typename Header>
+        PUGI__FN void node_copy_string(String& dest, Header& header, uintptr_t header_mask, char_t* source, Header& source_header, xml_allocator* alloc)
+        {
+                assert(!dest && (header & header_mask) == 0);
+
+                if (source)
+                {
+                        if (alloc && (source_header & header_mask) == 0)
+                        {
+                                dest = source;
+
+                                // since strcpy_insitu can reuse document buffer memory we need to mark both source and dest as shared
+                                header |= xml_memory_page_contents_shared_mask;
+                                source_header |= xml_memory_page_contents_shared_mask;
+                        }
+                        else
+                                strcpy_insitu(dest, header, header_mask, source, strlength(source));
+                }
+        }
+
+        PUGI__FN void node_copy_contents(xml_node_struct* dn, xml_node_struct* sn, xml_allocator* shared_alloc)
+        {
+                node_copy_string(dn->name, dn->header, xml_memory_page_name_allocated_mask, sn->name, sn->header, shared_alloc);
+                node_copy_string(dn->value, dn->header, xml_memory_page_value_allocated_mask, sn->value, sn->header, shared_alloc);
+
+                for (xml_attribute_struct* sa = sn->first_attribute; sa; sa = sa->next_attribute)
+                {
+                        xml_attribute_struct* da = append_new_attribute(dn, get_allocator(dn));
+
+                        if (da)
+                        {
+                                node_copy_string(da->name, da->header, xml_memory_page_name_allocated_mask, sa->name, sa->header, shared_alloc);
+                                node_copy_string(da->value, da->header, xml_memory_page_value_allocated_mask, sa->value, sa->header, shared_alloc);
+                        }
+                }
+        }
+
+        PUGI__FN void node_copy_tree(xml_node_struct* dn, xml_node_struct* sn)
+        {
+                xml_allocator& alloc = get_allocator(dn);
+                xml_allocator* shared_alloc = (&alloc == &get_allocator(sn)) ? &alloc : 0;
+
+                node_copy_contents(dn, sn, shared_alloc);
+
+                xml_node_struct* dit = dn;
+                xml_node_struct* sit = sn->first_child;
+
+                while (sit && sit != sn)
+                {
+                        if (sit != dn)
+                        {
+                                xml_node_struct* copy = append_new_node(dit, alloc, PUGI__NODETYPE(sit));
+
+                                if (copy)
+                                {
+                                        node_copy_contents(copy, sit, shared_alloc);
+
+                                        if (sit->first_child)
+                                        {
+                                                dit = copy;
+                                                sit = sit->first_child;
+                                                continue;
+                                        }
+                                }
+                        }
+
+                        // continue to the next node
+                        do
+                        {
+                                if (sit->next_sibling)
+                                {
+                                        sit = sit->next_sibling;
+                                        break;
+                                }
+
+                                sit = sit->parent;
+                                dit = dit->parent;
+                        }
+                        while (sit != sn);
+                }
+        }
+
+        PUGI__FN void node_copy_attribute(xml_attribute_struct* da, xml_attribute_struct* sa)
+        {
+                xml_allocator& alloc = get_allocator(da);
+                xml_allocator* shared_alloc = (&alloc == &get_allocator(sa)) ? &alloc : 0;
+
+                node_copy_string(da->name, da->header, xml_memory_page_name_allocated_mask, sa->name, sa->header, shared_alloc);
+                node_copy_string(da->value, da->header, xml_memory_page_value_allocated_mask, sa->value, sa->header, shared_alloc);
         }
 
         inline bool is_text_node(xml_node_struct* node)
         {
-                xml_node_type type = static_cast<xml_node_type>((node->header & impl::xml_memory_page_type_mask) + 1);
+                xml_node_type type = PUGI__NODETYPE(node);
 
                 return type == node_pcdata || type == node_cdata;
@@ -3289 +4337 @@
 
         // get value with conversion functions
-        PUGI__FN int get_value_int(const char_t* value, int def)
-        {
-                if (!value) return def;
-
-        #ifdef PUGIXML_WCHAR_MODE
-                return static_cast<int>(wcstol(value, 0, 10));
-        #else
-                return static_cast<int>(strtol(value, 0, 10));
-        #endif
-        }
-
-        PUGI__FN unsigned int get_value_uint(const char_t* value, unsigned int def)
-        {
-                if (!value) return def;
-
-        #ifdef PUGIXML_WCHAR_MODE
-                return static_cast<unsigned int>(wcstoul(value, 0, 10));
-        #else
-                return static_cast<unsigned int>(strtoul(value, 0, 10));
-        #endif
-        }
-
-        PUGI__FN double get_value_double(const char_t* value, double def)
-        {
-                if (!value) return def;
-
+        template <typename U> U string_to_integer(const char_t* value, U minneg, U maxpos)
+        {
+                U result = 0;
+                const char_t* s = value;
+
+                while (PUGI__IS_CHARTYPE(*s, ct_space))
+                        s++;
+
+                bool negative = (*s == '-');
+
+                s += (*s == '+' || *s == '-');
+
+                bool overflow = false;
+
+                if (s[0] == '0' && (s[1] | ' ') == 'x')
+                {
+                        s += 2;
+
+                        // since overflow detection relies on length of the sequence skip leading zeros
+                        while (*s == '0')
+                                s++;
+
+                        const char_t* start = s;
+
+                        for (;;)
+                        {
+                                if (static_cast<unsigned>(*s - '0') < 10)
+                                        result = result * 16 + (*s - '0');
+                                else if (static_cast<unsigned>((*s | ' ') - 'a') < 6)
+                                        result = result * 16 + ((*s | ' ') - 'a' + 10);
+                                else
+                                        break;
+
+                                s++;
+                        }
+
+                        size_t digits = static_cast<size_t>(s - start);
+
+                        overflow = digits > sizeof(U) * 2;
+                }
+                else
+                {
+                        // since overflow detection relies on length of the sequence skip leading zeros
+                        while (*s == '0')
+                                s++;
+
+                        const char_t* start = s;
+
+                        for (;;)
+                        {
+                                if (static_cast<unsigned>(*s - '0') < 10)
+                                        result = result * 10 + (*s - '0');
+                                else
+                                        break;
+
+                                s++;
+                        }
+
+                        size_t digits = static_cast<size_t>(s - start);
+
+                        PUGI__STATIC_ASSERT(sizeof(U) == 8 || sizeof(U) == 4 || sizeof(U) == 2);
+
+                        const size_t max_digits10 = sizeof(U) == 8 ? 20 : sizeof(U) == 4 ? 10 : 5;
+                        const char_t max_lead = sizeof(U) == 8 ? '1' : sizeof(U) == 4 ? '4' : '6';
+                        const size_t high_bit = sizeof(U) * 8 - 1;
+
+                        overflow = digits >= max_digits10 && !(digits == max_digits10 && (*start < max_lead || (*start == max_lead && result >> high_bit)));
+                }
+
+                if (negative)
+                        return (overflow || result > minneg) ? 0 - minneg : 0 - result;
+                else
+                        return (overflow || result > maxpos) ? maxpos : result;
+        }
+
+        PUGI__FN int get_value_int(const char_t* value)
+        {
+                return string_to_integer<unsigned int>(value, 0 - static_cast<unsigned int>(INT_MIN), INT_MAX);
+        }
+
+        PUGI__FN unsigned int get_value_uint(const char_t* value)
+        {
+                return string_to_integer<unsigned int>(value, 0, UINT_MAX);
+        }
+
+        PUGI__FN double get_value_double(const char_t* value)
+        {
         #ifdef PUGIXML_WCHAR_MODE
                 return wcstod(value, 0);
@@ -3322 +4431 @@
         }
 
-        PUGI__FN float get_value_float(const char_t* value, float def)
-        {
-                if (!value) return def;
-
+        PUGI__FN float get_value_float(const char_t* value)
+        {
         #ifdef PUGIXML_WCHAR_MODE
                 return static_cast<float>(wcstod(value, 0));
@@ -3333 +4440 @@
         }
 
-        PUGI__FN bool get_value_bool(const char_t* value, bool def)
-        {
-                if (!value) return def;
-
+        PUGI__FN bool get_value_bool(const char_t* value)
+        {
                 // only look at first char
                 char_t first = *value;
@@ -3344 +4449 @@
         }
 
+#ifdef PUGIXML_HAS_LONG_LONG
+        PUGI__FN long long get_value_llong(const char_t* value)
+        {
+                return string_to_integer<unsigned long long>(value, 0 - static_cast<unsigned long long>(LLONG_MIN), LLONG_MAX);
+        }
+
+        PUGI__FN unsigned long long get_value_ullong(const char_t* value)
+        {
+                return string_to_integer<unsigned long long>(value, 0, ULLONG_MAX);
+        }
+#endif
+
+        template <typename U>
+        PUGI__FN char_t* integer_to_string(char_t* begin, char_t* end, U value, bool negative)
+        {
+                char_t* result = end - 1;
+                U rest = negative ? 0 - value : value;
+
+                do
+                {
+                        *result-- = static_cast<char_t>('0' + (rest % 10));
+                        rest /= 10;
+                }
+                while (rest);
+
+                assert(result >= begin);
+                (void)begin;
+
+                *result = '-';
+
+                return result + !negative;
+        }
+
         // set value with conversion functions
-        PUGI__FN bool set_value_buffer(char_t*& dest, uintptr_t& header, uintptr_t header_mask, char (&buf)[128])
+        template <typename String, typename Header>
+        PUGI__FN bool set_value_ascii(String& dest, Header& header, uintptr_t header_mask, char* buf)
         {
         #ifdef PUGIXML_WCHAR_MODE
                 char_t wbuf[128];
-                impl::widen_ascii(wbuf, buf);
-
-                return strcpy_insitu(dest, header, header_mask, wbuf);
+                assert(strlen(buf) < sizeof(wbuf) / sizeof(wbuf[0]));
+
+                size_t offset = 0;
+                for (; buf[offset]; ++offset) wbuf[offset] = buf[offset];
+
+                return strcpy_insitu(dest, header, header_mask, wbuf, offset);
         #else
-                return strcpy_insitu(dest, header, header_mask, buf);
+                return strcpy_insitu(dest, header, header_mask, buf, strlen(buf));
         #endif
         }
 
-        PUGI__FN bool set_value_convert(char_t*& dest, uintptr_t& header, uintptr_t header_mask, int value)
+        template <typename String, typename Header>
+        PUGI__FN bool set_value_convert(String& dest, Header& header, uintptr_t header_mask, int value)
+        {
+                char_t buf[64];
+                char_t* end = buf + sizeof(buf) / sizeof(buf[0]);
+                char_t* begin = integer_to_string<unsigned int>(buf, end, value, value < 0);
+
+                return strcpy_insitu(dest, header, header_mask, begin, end - begin);
+        }
+
+        template <typename String, typename Header>
+        PUGI__FN bool set_value_convert(String& dest, Header& header, uintptr_t header_mask, unsigned int value)
+        {
+                char_t buf[64];
+                char_t* end = buf + sizeof(buf) / sizeof(buf[0]);
+                char_t* begin = integer_to_string<unsigned int>(buf, end, value, false);
+
+                return strcpy_insitu(dest, header, header_mask, begin, end - begin);
+        }
+
+        template <typename String, typename Header>
+        PUGI__FN bool set_value_convert(String& dest, Header& header, uintptr_t header_mask, float value)
         {
                 char buf[128];
-                sprintf(buf, "%d", value);
+                sprintf(buf, "%.9g", value);
+
+                return set_value_ascii(dest, header, header_mask, buf);
+        }
+
+        template <typename String, typename Header>
+        PUGI__FN bool set_value_convert(String& dest, Header& header, uintptr_t header_mask, double value)
+        {
+                char buf[128];
+                sprintf(buf, "%.17g", value);
+
+                return set_value_ascii(dest, header, header_mask, buf);
+        }
         
-                return set_value_buffer(dest, header, header_mask, buf);
-        }
-
-        PUGI__FN bool set_value_convert(char_t*& dest, uintptr_t& header, uintptr_t header_mask, unsigned int value)
-        {
-                char buf[128];
-                sprintf(buf, "%u", value);
-
-                return set_value_buffer(dest, header, header_mask, buf);
-        }
-
-        PUGI__FN bool set_value_convert(char_t*& dest, uintptr_t& header, uintptr_t header_mask, double value)
-        {
-                char buf[128];
-                sprintf(buf, "%g", value);
-
-                return set_value_buffer(dest, header, header_mask, buf);
-        }
-        
-        PUGI__FN bool set_value_convert(char_t*& dest, uintptr_t& header, uintptr_t header_mask, bool value)
-        {
-                return strcpy_insitu(dest, header, header_mask, value ? PUGIXML_TEXT("true") : PUGIXML_TEXT("false"));
+        template <typename String, typename Header>
+        PUGI__FN bool set_value_convert(String& dest, Header& header, uintptr_t header_mask, bool value)
+        {
+                return strcpy_insitu(dest, header, header_mask, value ? PUGIXML_TEXT("true") : PUGIXML_TEXT("false"), value ? 4 : 5);
+        }
+
+#ifdef PUGIXML_HAS_LONG_LONG
+        template <typename String, typename Header>
+        PUGI__FN bool set_value_convert(String& dest, Header& header, uintptr_t header_mask, long long value)
+        {
+                char_t buf[64];
+                char_t* end = buf + sizeof(buf) / sizeof(buf[0]);
+                char_t* begin = integer_to_string<unsigned long long>(buf, end, value, value < 0);
+
+                return strcpy_insitu(dest, header, header_mask, begin, end - begin);
+        }
+
+        template <typename String, typename Header>
+        PUGI__FN bool set_value_convert(String& dest, Header& header, uintptr_t header_mask, unsigned long long value)
+        {
+                char_t buf[64];
+                char_t* end = buf + sizeof(buf) / sizeof(buf[0]);
+                char_t* begin = integer_to_string<unsigned long long>(buf, end, value, false);
+
+                return strcpy_insitu(dest, header, header_mask, begin, end - begin);
+        }
+#endif
+
+        PUGI__FN xml_parse_result load_buffer_impl(xml_document_struct* doc, xml_node_struct* root, void* contents, size_t size, unsigned int options, xml_encoding encoding, bool is_mutable, bool own, char_t** out_buffer)
+        {
+                // check input buffer
+                if (!contents && size) return make_parse_result(status_io_error);
+
+                // get actual encoding
+                xml_encoding buffer_encoding = impl::get_buffer_encoding(encoding, contents, size);
+
+                // get private buffer
+                char_t* buffer = 0;
+                size_t length = 0;
+
+                if (!impl::convert_buffer(buffer, length, buffer_encoding, contents, size, is_mutable)) return impl::make_parse_result(status_out_of_memory);
+
+                // delete original buffer if we performed a conversion
+                if (own && buffer != contents && contents) impl::xml_memory::deallocate(contents);
+
+                // grab onto buffer if it's our buffer, user is responsible for deallocating contents himself
+                if (own || buffer != contents) *out_buffer = buffer;
+
+                // store buffer for offset_debug
+                doc->buffer = buffer;
+
+                // parse
+                xml_parse_result res = impl::xml_parser::parse(buffer, length, doc, root, options);
+
+                // remember encoding
+                res.encoding = buffer_encoding;
+
+                return res;
         }
 
@@ -3396 +4607 @@
                 length_type length = _ftelli64(file);
                 _fseeki64(file, 0, SEEK_SET);
-        #elif defined(__MINGW32__) && !defined(__NO_MINGW_LFS) && !defined(__STRICT_ANSI__)
+        #elif defined(__MINGW32__) && !defined(__NO_MINGW_LFS) && (!defined(__STRICT_ANSI__) || defined(__MINGW64_VERSION_MAJOR))
                 // there are 64-bit versions of fseek/ftell, let's use them
                 typedef off64_t length_type;
@@ -3426 +4637 @@
         }
 
-        PUGI__FN xml_parse_result load_file_impl(xml_document& doc, FILE* file, unsigned int options, xml_encoding encoding)
+        // This function assumes that buffer has extra sizeof(char_t) writable bytes after size
+        PUGI__FN size_t zero_terminate_buffer(void* buffer, size_t size, xml_encoding encoding) 
+        {
+                // We only need to zero-terminate if encoding conversion does not do it for us
+        #ifdef PUGIXML_WCHAR_MODE
+                xml_encoding wchar_encoding = get_wchar_encoding();
+
+                if (encoding == wchar_encoding || need_endian_swap_utf(encoding, wchar_encoding))
+                {
+                        size_t length = size / sizeof(char_t);
+
+                        static_cast<char_t*>(buffer)[length] = 0;
+                        return (length + 1) * sizeof(char_t);
+                }
+        #else
+                if (encoding == encoding_utf8)
+                {
+                        static_cast<char*>(buffer)[size] = 0;
+                        return size + 1;
+                }
+        #endif
+
+                return size;
+        }
+
+        PUGI__FN xml_parse_result load_file_impl(xml_document_struct* doc, FILE* file, unsigned int options, xml_encoding encoding, char_t** out_buffer)
         {
                 if (!file) return make_parse_result(status_file_not_found);
@@ -3433 +4669 @@
                 size_t size = 0;
                 xml_parse_status size_status = get_file_size(file, size);
-
-                if (size_status != status_ok)
-                {
-                        fclose(file);
-                        return make_parse_result(size_status);
-                }
+                if (size_status != status_ok) return make_parse_result(size_status);
                 
+                size_t max_suffix_size = sizeof(char_t);
+
                 // allocate buffer for the whole file
-                char* contents = static_cast<char*>(xml_memory::allocate(size > 0 ? size : 1));
-
-                if (!contents)
-                {
-                        fclose(file);
-                        return make_parse_result(status_out_of_memory);
-                }
+                char* contents = static_cast<char*>(xml_memory::allocate(size + max_suffix_size));
+                if (!contents) return make_parse_result(status_out_of_memory);
 
                 // read file in memory
                 size_t read_size = fread(contents, 1, size, file);
-                fclose(file);
 
                 if (read_size != size)
@@ -3458 +4685 @@
                         return make_parse_result(status_io_error);
                 }
-                
-                return doc.load_buffer_inplace_own(contents, size, options, encoding);
+
+                xml_encoding real_encoding = get_buffer_encoding(encoding, contents, size);
+
+                return load_buffer_impl(doc, doc, contents, zero_terminate_buffer(contents, size, real_encoding), options, real_encoding, true, true, out_buffer);
         }
 
@@ -3468 +4697 @@
                 {
                         void* memory = xml_memory::allocate(sizeof(xml_stream_chunk));
+                        if (!memory) return 0;
                         
                         return new (memory) xml_stream_chunk();
                 }
 
-                static void destroy(void* ptr)
-                {
-                        xml_stream_chunk* chunk = static_cast<xml_stream_chunk*>(ptr);
-
+                static void destroy(xml_stream_chunk* chunk)
+                {
                         // free chunk chain
                         while (chunk)
                         {
-                                xml_stream_chunk* next = chunk->next;
+                                xml_stream_chunk* next_ = chunk->next;
+
                                 xml_memory::deallocate(chunk);
-                                chunk = next;
+
+                                chunk = next_;
                         }
                 }
@@ -3497 +4727 @@
         template <typename T> PUGI__FN xml_parse_status load_stream_data_noseek(std::basic_istream<T>& stream, void** out_buffer, size_t* out_size)
         {
-                buffer_holder chunks(0, xml_stream_chunk<T>::destroy);
+                auto_deleter<xml_stream_chunk<T> > chunks(0, xml_stream_chunk<T>::destroy);
 
                 // read file to a chunk list
@@ -3525 +4755 @@
                 }
 
+                size_t max_suffix_size = sizeof(char_t);
+
                 // copy chunk list to a contiguous buffer
-                char* buffer = static_cast<char*>(xml_memory::allocate(total));
+                char* buffer = static_cast<char*>(xml_memory::allocate(total + max_suffix_size));
                 if (!buffer) return status_out_of_memory;
 
                 char* write = buffer;
 
-                for (xml_stream_chunk<T>* chunk = static_cast<xml_stream_chunk<T>*>(chunks.data); chunk; chunk = chunk->next)
+                for (xml_stream_chunk<T>* chunk = chunks.data; chunk; chunk = chunk->next)
                 {
                         assert(write + chunk->size <= buffer + total);
@@ -3562 +4794 @@
                 if (static_cast<std::streamsize>(read_length) != length || length < 0) return status_out_of_memory;
 
+                size_t max_suffix_size = sizeof(char_t);
+
                 // read stream data into memory (guard against stream exceptions with buffer holder)
-                buffer_holder buffer(xml_memory::allocate((read_length > 0 ? read_length : 1) * sizeof(T)), xml_memory::deallocate);
+                auto_deleter<void> buffer(xml_memory::allocate(read_length * sizeof(T) + max_suffix_size), xml_memory::deallocate);
                 if (!buffer.data) return status_out_of_memory;
 
@@ -3574 +4808 @@
                 size_t actual_length = static_cast<size_t>(stream.gcount());
                 assert(actual_length <= read_length);
-
+                
                 *out_buffer = buffer.release();
                 *out_size = actual_length * sizeof(T);
@@ -3581 +4815 @@
         }
 
-        template <typename T> PUGI__FN xml_parse_result load_stream_impl(xml_document& doc, std::basic_istream<T>& stream, unsigned int options, xml_encoding encoding)
+        template <typename T> PUGI__FN xml_parse_result load_stream_impl(xml_document_struct* doc, std::basic_istream<T>& stream, unsigned int options, xml_encoding encoding, char_t** out_buffer)
         {
                 void* buffer = 0;
                 size_t size = 0;
+                xml_parse_status status = status_ok;
+
+                // if stream has an error bit set, bail out (otherwise tellg() can fail and we'll clear error bits)
+                if (stream.fail()) return make_parse_result(status_io_error);
 
                 // load stream to memory (using seek-based implementation if possible, since it's faster and takes less memory)
-                xml_parse_status status = (stream.tellg() < 0) ? load_stream_data_noseek(stream, &buffer, &size) : load_stream_data_seek(stream, &buffer, &size);
+                if (stream.tellg() < 0)
+                {
+                        stream.clear(); // clear error flags that could be set by a failing tellg
+                        status = load_stream_data_noseek(stream, &buffer, &size);
+                }
+                else
+                        status = load_stream_data_seek(stream, &buffer, &size);
+
                 if (status != status_ok) return make_parse_result(status);
 
-                return doc.load_buffer_inplace_own(buffer, size, options, encoding);
+                xml_encoding real_encoding = get_buffer_encoding(encoding, buffer, size);
+                
+                return load_buffer_impl(doc, doc, buffer, zero_terminate_buffer(buffer, size, real_encoding), options, real_encoding, true, true, out_buffer);
         }
 #endif
 
-#if defined(PUGI__MSVC_CRT_VERSION) || defined(__BORLANDC__) || (defined(__MINGW32__) && !defined(__STRICT_ANSI__))
+#if defined(PUGI__MSVC_CRT_VERSION) || defined(__BORLANDC__) || (defined(__MINGW32__) && (!defined(__STRICT_ANSI__) || defined(__MINGW64_VERSION_MAJOR)))
         PUGI__FN FILE* open_file_wide(const wchar_t* path, const wchar_t* mode)
         {
@@ -3605 +4852 @@
 
                 // first pass: get length in utf8 characters
-                size_t length = wcslen(str);
+                size_t length = strlength_wide(str);
                 size_t size = as_utf8_begin(str, length);
 
@@ -3615 +4862 @@
                 as_utf8_end(result, size, str, length);
 
+                // zero-terminate
+                result[size] = 0;
+
                 return result;
         }
@@ -3645 +4895 @@
                 doc.save(writer, indent, flags, encoding);
 
-                int result = ferror(file);
-
-                fclose(file);
-
-                return result == 0;
-        }
+                return ferror(file) == 0;
+        }
+
+        struct name_null_sentry
+        {
+                xml_node_struct* node;
+                char_t* name;
+
+                name_null_sentry(xml_node_struct* node_): node(node_), name(node_->name)
+                {
+                        node->name = 0;
+                }
+
+                ~name_null_sentry()
+                {
+                        node->name = name;
+                }
+        };
 PUGI__NS_END
 
@@ -3778 +5040 @@
         PUGI__FN const char_t* xml_attribute::as_string(const char_t* def) const
         {
-                return (_attr && _attr->value) ? _attr->value : def;
+                return (_attr && _attr->value) ? _attr->value + 0 : def;
         }
 
         PUGI__FN int xml_attribute::as_int(int def) const
         {
-                return impl::get_value_int(_attr ? _attr->value : 0, def);
+                return (_attr && _attr->value) ? impl::get_value_int(_attr->value) : def;
         }
 
         PUGI__FN unsigned int xml_attribute::as_uint(unsigned int def) const
         {
-                return impl::get_value_uint(_attr ? _attr->value : 0, def);
+                return (_attr && _attr->value) ? impl::get_value_uint(_attr->value) : def;
         }
 
         PUGI__FN double xml_attribute::as_double(double def) const
         {
-                return impl::get_value_double(_attr ? _attr->value : 0, def);
+                return (_attr && _attr->value) ? impl::get_value_double(_attr->value) : def;
         }
 
         PUGI__FN float xml_attribute::as_float(float def) const
         {
-                return impl::get_value_float(_attr ? _attr->value : 0, def);
+                return (_attr && _attr->value) ? impl::get_value_float(_attr->value) : def;
         }
 
         PUGI__FN bool xml_attribute::as_bool(bool def) const
         {
-                return impl::get_value_bool(_attr ? _attr->value : 0, def);
-        }
+                return (_attr && _attr->value) ? impl::get_value_bool(_attr->value) : def;
+        }
+
+#ifdef PUGIXML_HAS_LONG_LONG
+        PUGI__FN long long xml_attribute::as_llong(long long def) const
+        {
+                return (_attr && _attr->value) ? impl::get_value_llong(_attr->value) : def;
+        }
+
+        PUGI__FN unsigned long long xml_attribute::as_ullong(unsigned long long def) const
+        {
+                return (_attr && _attr->value) ? impl::get_value_ullong(_attr->value) : def;
+        }
+#endif
 
         PUGI__FN bool xml_attribute::empty() const
@@ -3813 +5087 @@
         PUGI__FN const char_t* xml_attribute::name() const
         {
-                return (_attr && _attr->name) ? _attr->name : PUGIXML_TEXT("");
+                return (_attr && _attr->name) ? _attr->name + 0 : PUGIXML_TEXT("");
         }
 
         PUGI__FN const char_t* xml_attribute::value() const
         {
-                return (_attr && _attr->value) ? _attr->value : PUGIXML_TEXT("");
+                return (_attr && _attr->value) ? _attr->value + 0 : PUGIXML_TEXT("");
         }
 
@@ -3855 +5129 @@
         }
         
-        PUGI__FN xml_attribute& xml_attribute::operator=(bool rhs)
+        PUGI__FN xml_attribute& xml_attribute::operator=(float rhs)
         {
                 set_value(rhs);
@@ -3861 +5135 @@
         }
 
+        PUGI__FN xml_attribute& xml_attribute::operator=(bool rhs)
+        {
+                set_value(rhs);
+                return *this;
+        }
+
+#ifdef PUGIXML_HAS_LONG_LONG
+        PUGI__FN xml_attribute& xml_attribute::operator=(long long rhs)
+        {
+                set_value(rhs);
+                return *this;
+        }
+
+        PUGI__FN xml_attribute& xml_attribute::operator=(unsigned long long rhs)
+        {
+                set_value(rhs);
+                return *this;
+        }
+#endif
+
         PUGI__FN bool xml_attribute::set_name(const char_t* rhs)
         {
                 if (!_attr) return false;
                 
-                return impl::strcpy_insitu(_attr->name, _attr->header, impl::xml_memory_page_name_allocated_mask, rhs);
+                return impl::strcpy_insitu(_attr->name, _attr->header, impl::xml_memory_page_name_allocated_mask, rhs, impl::strlength(rhs));
         }
                 
@@ -3872 +5166 @@
                 if (!_attr) return false;
 
-                return impl::strcpy_insitu(_attr->value, _attr->header, impl::xml_memory_page_value_allocated_mask, rhs);
+                return impl::strcpy_insitu(_attr->value, _attr->header, impl::xml_memory_page_value_allocated_mask, rhs, impl::strlength(rhs));
         }
 
@@ -3896 +5190 @@
         }
         
+        PUGI__FN bool xml_attribute::set_value(float rhs)
+        {
+                if (!_attr) return false;
+
+                return impl::set_value_convert(_attr->value, _attr->header, impl::xml_memory_page_value_allocated_mask, rhs);
+        }
+
         PUGI__FN bool xml_attribute::set_value(bool rhs)
         {
@@ -3902 +5203 @@
                 return impl::set_value_convert(_attr->value, _attr->header, impl::xml_memory_page_value_allocated_mask, rhs);
         }
+
+#ifdef PUGIXML_HAS_LONG_LONG
+        PUGI__FN bool xml_attribute::set_value(long long rhs)
+        {
+                if (!_attr) return false;
+
+                return impl::set_value_convert(_attr->value, _attr->header, impl::xml_memory_page_value_allocated_mask, rhs);
+        }
+
+        PUGI__FN bool xml_attribute::set_value(unsigned long long rhs)
+        {
+                if (!_attr) return false;
+
+                return impl::set_value_convert(_attr->value, _attr->header, impl::xml_memory_page_value_allocated_mask, rhs);
+        }
+#endif
 
 #ifdef __BORLANDC__
@@ -3939 +5256 @@
         PUGI__FN xml_node::iterator xml_node::begin() const
         {
-                return iterator(_root ? _root->first_child : 0, _root);
+                return iterator(_root ? _root->first_child + 0 : 0, _root);
         }
 
@@ -3949 +5266 @@
         PUGI__FN xml_node::attribute_iterator xml_node::attributes_begin() const
         {
-                return attribute_iterator(_root ? _root->first_attribute : 0, _root);
+                return attribute_iterator(_root ? _root->first_attribute + 0 : 0, _root);
         }
 
@@ -3964 +5281 @@
         PUGI__FN xml_object_range<xml_named_node_iterator> xml_node::children(const char_t* name_) const
         {
-                return xml_object_range<xml_named_node_iterator>(xml_named_node_iterator(child(name_), name_), xml_named_node_iterator());
+                return xml_object_range<xml_named_node_iterator>(xml_named_node_iterator(child(name_)._root, _root, name_), xml_named_node_iterator(0, _root, name_));
         }
 
@@ -4009 +5326 @@
         PUGI__FN const char_t* xml_node::name() const
         {
-                return (_root && _root->name) ? _root->name : PUGIXML_TEXT("");
+                return (_root && _root->name) ? _root->name + 0 : PUGIXML_TEXT("");
         }
 
         PUGI__FN xml_node_type xml_node::type() const
         {
-                return _root ? static_cast<xml_node_type>((_root->header & impl::xml_memory_page_type_mask) + 1) : node_null;
+                return _root ? PUGI__NODETYPE(_root) : node_null;
         }
         
         PUGI__FN const char_t* xml_node::value() const
         {
-                return (_root && _root->value) ? _root->value : PUGIXML_TEXT("");
+                return (_root && _root->value) ? _root->value + 0 : PUGIXML_TEXT("");
         }
         
@@ -4055 +5372 @@
         PUGI__FN xml_node xml_node::next_sibling() const
         {
-                if (!_root) return xml_node();
-                
-                if (_root->next_sibling) return xml_node(_root->next_sibling);
-                else return xml_node();
+                return _root ? xml_node(_root->next_sibling) : xml_node();
         }
 
@@ -4071 +5385 @@
         }
 
+        PUGI__FN xml_attribute xml_node::attribute(const char_t* name_, xml_attribute& hint_) const
+        {
+                xml_attribute_struct* hint = hint_._attr;
+
+                // if hint is not an attribute of node, behavior is not defined
+                assert(!hint || (_root && impl::is_attribute_of(hint, _root)));
+
+                if (!_root) return xml_attribute();
+
+                // optimistically search from hint up until the end
+                for (xml_attribute_struct* i = hint; i; i = i->next_attribute)
+                        if (i->name && impl::strequal(name_, i->name))
+                        {
+                                // update hint to maximize efficiency of searching for consecutive attributes
+                                hint_._attr = i->next_attribute;
+
+                                return xml_attribute(i);
+                        }
+
+                // wrap around and search from the first attribute until the hint
+                // 'j' null pointer check is technically redundant, but it prevents a crash in case the assertion above fails
+                for (xml_attribute_struct* j = _root->first_attribute; j && j != hint; j = j->next_attribute)
+                        if (j->name && impl::strequal(name_, j->name))
+                        {
+                                // update hint to maximize efficiency of searching for consecutive attributes
+                                hint_._attr = j->next_attribute;
+
+                                return xml_attribute(j);
+                        }
+
+                return xml_attribute();
+        }
+
         PUGI__FN xml_node xml_node::previous_sibling() const
         {
@@ -4086 +5433 @@
         PUGI__FN xml_node xml_node::root() const
         {
-                if (!_root) return xml_node();
-
-                impl::xml_memory_page* page = reinterpret_cast<impl::xml_memory_page*>(_root->header & impl::xml_memory_page_pointer_mask);
-
-                return xml_node(static_cast<impl::xml_document_struct*>(page->allocator));
+                return _root ? xml_node(&impl::get_document(_root)) : xml_node();
         }
 
@@ -4103 +5446 @@
                 
                 for (xml_node_struct* i = _root->first_child; i; i = i->next_sibling)
-                        if (i->value && impl::is_text_node(i))
+                        if (impl::is_text_node(i) && i->value)
                                 return i->value;
 
@@ -4136 +5479 @@
         PUGI__FN bool xml_node::set_name(const char_t* rhs)
         {
-                switch (type())
-                {
-                case node_pi:
-                case node_declaration:
-                case node_element:
-                        return impl::strcpy_insitu(_root->name, _root->header, impl::xml_memory_page_name_allocated_mask, rhs);
-
-                default:
+                xml_node_type type_ = _root ? PUGI__NODETYPE(_root) : node_null;
+
+                if (type_ != node_element && type_ != node_pi && type_ != node_declaration)
                         return false;
-                }
+
+                return impl::strcpy_insitu(_root->name, _root->header, impl::xml_memory_page_name_allocated_mask, rhs, impl::strlength(rhs));
         }
                 
         PUGI__FN bool xml_node::set_value(const char_t* rhs)
         {
-                switch (type())
-                {
-                case node_pi:
-                case node_cdata:
-                case node_pcdata:
-                case node_comment:
-                case node_doctype:
-                        return impl::strcpy_insitu(_root->value, _root->header, impl::xml_memory_page_value_allocated_mask, rhs);
-
-                default:
+                xml_node_type type_ = _root ? PUGI__NODETYPE(_root) : node_null;
+
+                if (type_ != node_pcdata && type_ != node_cdata && type_ != node_comment && type_ != node_pi && type_ != node_doctype)
                         return false;
-                }
+
+                return impl::strcpy_insitu(_root->value, _root->header, impl::xml_memory_page_value_allocated_mask, rhs, impl::strlength(rhs));
         }
 
         PUGI__FN xml_attribute xml_node::append_attribute(const char_t* name_)
         {
-                if (type() != node_element && type() != node_declaration) return xml_attribute();
+                if (!impl::allow_insert_attribute(type())) return xml_attribute();
                 
-                xml_attribute a(impl::append_attribute_ll(_root, impl::get_allocator(_root)));
+                impl::xml_allocator& alloc = impl::get_allocator(_root);
+                if (!alloc.reserve()) return xml_attribute();
+
+                xml_attribute a(impl::allocate_attribute(alloc));
+                if (!a) return xml_attribute();
+
+                impl::append_attribute(a._attr, _root);
+
                 a.set_name(name_);
                 
@@ -4176 +5516 @@
         PUGI__FN xml_attribute xml_node::prepend_attribute(const char_t* name_)
         {
-                if (type() != node_element && type() != node_declaration) return xml_attribute();
+                if (!impl::allow_insert_attribute(type())) return xml_attribute();
                 
-                xml_attribute a(impl::allocate_attribute(impl::get_allocator(_root)));
+                impl::xml_allocator& alloc = impl::get_allocator(_root);
+                if (!alloc.reserve()) return xml_attribute();
+
+                xml_attribute a(impl::allocate_attribute(alloc));
                 if (!a) return xml_attribute();
 
+                impl::prepend_attribute(a._attr, _root);
+
                 a.set_name(name_);
+
+                return a;
+        }
+
+        PUGI__FN xml_attribute xml_node::insert_attribute_after(const char_t* name_, const xml_attribute& attr)
+        {
+                if (!impl::allow_insert_attribute(type())) return xml_attribute();
+                if (!attr || !impl::is_attribute_of(attr._attr, _root)) return xml_attribute();
                 
-                xml_attribute_struct* head = _root->first_attribute;
-
-                if (head)
-                {
-                        a._attr->prev_attribute_c = head->prev_attribute_c;
-                        head->prev_attribute_c = a._attr;
-                }
-                else
-                        a._attr->prev_attribute_c = a._attr;
+                impl::xml_allocator& alloc = impl::get_allocator(_root);
+                if (!alloc.reserve()) return xml_attribute();
+
+                xml_attribute a(impl::allocate_attribute(alloc));
+                if (!a) return xml_attribute();
+
+                impl::insert_attribute_after(a._attr, attr._attr, _root);
+
+                a.set_name(name_);
+
+                return a;
+        }
+
+        PUGI__FN xml_attribute xml_node::insert_attribute_before(const char_t* name_, const xml_attribute& attr)
+        {
+                if (!impl::allow_insert_attribute(type())) return xml_attribute();
+                if (!attr || !impl::is_attribute_of(attr._attr, _root)) return xml_attribute();
                 
-                a._attr->next_attribute = head;
-                _root->first_attribute = a._attr;
-                                
+                impl::xml_allocator& alloc = impl::get_allocator(_root);
+                if (!alloc.reserve()) return xml_attribute();
+
+                xml_attribute a(impl::allocate_attribute(alloc));
+                if (!a) return xml_attribute();
+
+                impl::insert_attribute_before(a._attr, attr._attr, _root);
+
+                a.set_name(name_);
+
                 return a;
         }
 
-        PUGI__FN xml_attribute xml_node::insert_attribute_before(const char_t* name_, const xml_attribute& attr)
-        {
-                if ((type() != node_element && type() != node_declaration) || attr.empty()) return xml_attribute();
+        PUGI__FN xml_attribute xml_node::append_copy(const xml_attribute& proto)
+        {
+                if (!proto) return xml_attribute();
+                if (!impl::allow_insert_attribute(type())) return xml_attribute();
+
+                impl::xml_allocator& alloc = impl::get_allocator(_root);
+                if (!alloc.reserve()) return xml_attribute();
+
+                xml_attribute a(impl::allocate_attribute(alloc));
+                if (!a) return xml_attribute();
+
+                impl::append_attribute(a._attr, _root);
+                impl::node_copy_attribute(a._attr, proto._attr);
+
+                return a;
+        }
+
+        PUGI__FN xml_attribute xml_node::prepend_copy(const xml_attribute& proto)
+        {
+                if (!proto) return xml_attribute();
+                if (!impl::allow_insert_attribute(type())) return xml_attribute();
+
+                impl::xml_allocator& alloc = impl::get_allocator(_root);
+                if (!alloc.reserve()) return xml_attribute();
+
+                xml_attribute a(impl::allocate_attribute(alloc));
+                if (!a) return xml_attribute();
+
+                impl::prepend_attribute(a._attr, _root);
+                impl::node_copy_attribute(a._attr, proto._attr);
+
+                return a;
+        }
+
+        PUGI__FN xml_attribute xml_node::insert_copy_after(const xml_attribute& proto, const xml_attribute& attr)
+        {
+                if (!proto) return xml_attribute();
+                if (!impl::allow_insert_attribute(type())) return xml_attribute();
+                if (!attr || !impl::is_attribute_of(attr._attr, _root)) return xml_attribute();
+
+                impl::xml_allocator& alloc = impl::get_allocator(_root);
+                if (!alloc.reserve()) return xml_attribute();
+
+                xml_attribute a(impl::allocate_attribute(alloc));
+                if (!a) return xml_attribute();
+
+                impl::insert_attribute_after(a._attr, attr._attr, _root);
+                impl::node_copy_attribute(a._attr, proto._attr);
+
+                return a;
+        }
+
+        PUGI__FN xml_attribute xml_node::insert_copy_before(const xml_attribute& proto, const xml_attribute& attr)
+        {
+                if (!proto) return xml_attribute();
+                if (!impl::allow_insert_attribute(type())) return xml_attribute();
+                if (!attr || !impl::is_attribute_of(attr._attr, _root)) return xml_attribute();
+
+                impl::xml_allocator& alloc = impl::get_allocator(_root);
+                if (!alloc.reserve()) return xml_attribute();
+
+                xml_attribute a(impl::allocate_attribute(alloc));
+                if (!a) return xml_attribute();
+
+                impl::insert_attribute_before(a._attr, attr._attr, _root);
+                impl::node_copy_attribute(a._attr, proto._attr);
+
+                return a;
+        }
+
+        PUGI__FN xml_node xml_node::append_child(xml_node_type type_)
+        {
+                if (!impl::allow_insert_child(type(), type_)) return xml_node();
                 
-                // check that attribute belongs to *this
-                xml_attribute_struct* cur = attr._attr;
-
-                while (cur->prev_attribute_c->next_attribute) cur = cur->prev_attribute_c;
-
-                if (cur != _root->first_attribute) return xml_attribute();
-
-                xml_attribute a(impl::allocate_attribute(impl::get_allocator(_root)));
-                if (!a) return xml_attribute();
-
-                a.set_name(name_);
-
-                if (attr._attr->prev_attribute_c->next_attribute)
-                        attr._attr->prev_attribute_c->next_attribute = a._attr;
-                else
-                        _root->first_attribute = a._attr;
+                impl::xml_allocator& alloc = impl::get_allocator(_root);
+                if (!alloc.reserve()) return xml_node();
+
+                xml_node n(impl::allocate_node(alloc, type_));
+                if (!n) return xml_node();
+
+                impl::append_node(n._root, _root);
+
+                if (type_ == node_declaration) n.set_name(PUGIXML_TEXT("xml"));
+
+                return n;
+        }
+
+        PUGI__FN xml_node xml_node::prepend_child(xml_node_type type_)
+        {
+                if (!impl::allow_insert_child(type(), type_)) return xml_node();
+
+                impl::xml_allocator& alloc = impl::get_allocator(_root);
+                if (!alloc.reserve()) return xml_node();
                 
-                a._attr->prev_attribute_c = attr._attr->prev_attribute_c;
-                a._attr->next_attribute = attr._attr;
-                attr._attr->prev_attribute_c = a._attr;
-                                
-                return a;
-        }
-
-        PUGI__FN xml_attribute xml_node::insert_attribute_after(const char_t* name_, const xml_attribute& attr)
-        {
-                if ((type() != node_element && type() != node_declaration) || attr.empty()) return xml_attribute();
-                
-                // check that attribute belongs to *this
-                xml_attribute_struct* cur = attr._attr;
-
-                while (cur->prev_attribute_c->next_attribute) cur = cur->prev_attribute_c;
-
-                if (cur != _root->first_attribute) return xml_attribute();
-
-                xml_attribute a(impl::allocate_attribute(impl::get_allocator(_root)));
-                if (!a) return xml_attribute();
-
-                a.set_name(name_);
-
-                if (attr._attr->next_attribute)
-                        attr._attr->next_attribute->prev_attribute_c = a._attr;
-                else
-                        _root->first_attribute->prev_attribute_c = a._attr;
-                
-                a._attr->next_attribute = attr._attr->next_attribute;
-                a._attr->prev_attribute_c = attr._attr;
-                attr._attr->next_attribute = a._attr;
-
-                return a;
-        }
-
-        PUGI__FN xml_attribute xml_node::append_copy(const xml_attribute& proto)
-        {
-                if (!proto) return xml_attribute();
-
-                xml_attribute result = append_attribute(proto.name());
-                result.set_value(proto.value());
-
-                return result;
-        }
-
-        PUGI__FN xml_attribute xml_node::prepend_copy(const xml_attribute& proto)
-        {
-                if (!proto) return xml_attribute();
-
-                xml_attribute result = prepend_attribute(proto.name());
-                result.set_value(proto.value());
-
-                return result;
-        }
-
-        PUGI__FN xml_attribute xml_node::insert_copy_after(const xml_attribute& proto, const xml_attribute& attr)
-        {
-                if (!proto) return xml_attribute();
-
-                xml_attribute result = insert_attribute_after(proto.name(), attr);
-                result.set_value(proto.value());
-
-                return result;
-        }
-
-        PUGI__FN xml_attribute xml_node::insert_copy_before(const xml_attribute& proto, const xml_attribute& attr)
-        {
-                if (!proto) return xml_attribute();
-
-                xml_attribute result = insert_attribute_before(proto.name(), attr);
-                result.set_value(proto.value());
-
-                return result;
-        }
-
-        PUGI__FN xml_node xml_node::append_child(xml_node_type type_)
-        {
-                if (!impl::allow_insert_child(this->type(), type_)) return xml_node();
-                
-                xml_node n(impl::append_node(_root, impl::get_allocator(_root), type_));
-
-                if (type_ == node_declaration) n.set_name(PUGIXML_TEXT("xml"));
-
-                return n;
-        }
-
-        PUGI__FN xml_node xml_node::prepend_child(xml_node_type type_)
-        {
-                if (!impl::allow_insert_child(this->type(), type_)) return xml_node();
-                
-                xml_node n(impl::allocate_node(impl::get_allocator(_root), type_));
+                xml_node n(impl::allocate_node(alloc, type_));
                 if (!n) return xml_node();
 
-                n._root->parent = _root;
-
-                xml_node_struct* head = _root->first_child;
-
-                if (head)
-                {
-                        n._root->prev_sibling_c = head->prev_sibling_c;
-                        head->prev_sibling_c = n._root;
-                }
-                else
-                        n._root->prev_sibling_c = n._root;
-                
-                n._root->next_sibling = head;
-                _root->first_child = n._root;
+                impl::prepend_node(n._root, _root);
                                 
                 if (type_ == node_declaration) n.set_name(PUGIXML_TEXT("xml"));
@@ -4335 +5673 @@
         PUGI__FN xml_node xml_node::insert_child_before(xml_node_type type_, const xml_node& node)
         {
-                if (!impl::allow_insert_child(this->type(), type_)) return xml_node();
+                if (!impl::allow_insert_child(type(), type_)) return xml_node();
                 if (!node._root || node._root->parent != _root) return xml_node();
+
+                impl::xml_allocator& alloc = impl::get_allocator(_root);
+                if (!alloc.reserve()) return xml_node();
         
-                xml_node n(impl::allocate_node(impl::get_allocator(_root), type_));
+                xml_node n(impl::allocate_node(alloc, type_));
                 if (!n) return xml_node();
 
-                n._root->parent = _root;
+                impl::insert_node_before(n._root, node._root);
+
+                if (type_ == node_declaration) n.set_name(PUGIXML_TEXT("xml"));
+
+                return n;
+        }
+
+        PUGI__FN xml_node xml_node::insert_child_after(xml_node_type type_, const xml_node& node)
+        {
+                if (!impl::allow_insert_child(type(), type_)) return xml_node();
+                if (!node._root || node._root->parent != _root) return xml_node();
+
+                impl::xml_allocator& alloc = impl::get_allocator(_root);
+                if (!alloc.reserve()) return xml_node();
+        
+                xml_node n(impl::allocate_node(alloc, type_));
+                if (!n) return xml_node();
+
+                impl::insert_node_after(n._root, node._root);
+
+                if (type_ == node_declaration) n.set_name(PUGIXML_TEXT("xml"));
+
+                return n;
+        }
+
+        PUGI__FN xml_node xml_node::append_child(const char_t* name_)
+        {
+                xml_node result = append_child(node_element);
+
+                result.set_name(name_);
+
+                return result;
+        }
+
+        PUGI__FN xml_node xml_node::prepend_child(const char_t* name_)
+        {
+                xml_node result = prepend_child(node_element);
+
+                result.set_name(name_);
+
+                return result;
+        }
+
+        PUGI__FN xml_node xml_node::insert_child_after(const char_t* name_, const xml_node& node)
+        {
+                xml_node result = insert_child_after(node_element, node);
+
+                result.set_name(name_);
+
+                return result;
+        }
+
+        PUGI__FN xml_node xml_node::insert_child_before(const char_t* name_, const xml_node& node)
+        {
+                xml_node result = insert_child_before(node_element, node);
+
+                result.set_name(name_);
+
+                return result;
+        }
+
+        PUGI__FN xml_node xml_node::append_copy(const xml_node& proto)
+        {
+                xml_node_type type_ = proto.type();
+                if (!impl::allow_insert_child(type(), type_)) return xml_node();
+
+                impl::xml_allocator& alloc = impl::get_allocator(_root);
+                if (!alloc.reserve()) return xml_node();
+
+                xml_node n(impl::allocate_node(alloc, type_));
+                if (!n) return xml_node();
+
+                impl::append_node(n._root, _root);
+                impl::node_copy_tree(n._root, proto._root);
+
+                return n;
+        }
+
+        PUGI__FN xml_node xml_node::prepend_copy(const xml_node& proto)
+        {
+                xml_node_type type_ = proto.type();
+                if (!impl::allow_insert_child(type(), type_)) return xml_node();
+
+                impl::xml_allocator& alloc = impl::get_allocator(_root);
+                if (!alloc.reserve()) return xml_node();
+
+                xml_node n(impl::allocate_node(alloc, type_));
+                if (!n) return xml_node();
+
+                impl::prepend_node(n._root, _root);
+                impl::node_copy_tree(n._root, proto._root);
+
+                return n;
+        }
+
+        PUGI__FN xml_node xml_node::insert_copy_after(const xml_node& proto, const xml_node& node)
+        {
+                xml_node_type type_ = proto.type();
+                if (!impl::allow_insert_child(type(), type_)) return xml_node();
+                if (!node._root || node._root->parent != _root) return xml_node();
+
+                impl::xml_allocator& alloc = impl::get_allocator(_root);
+                if (!alloc.reserve()) return xml_node();
+
+                xml_node n(impl::allocate_node(alloc, type_));
+                if (!n) return xml_node();
+
+                impl::insert_node_after(n._root, node._root);
+                impl::node_copy_tree(n._root, proto._root);
+
+                return n;
+        }
+
+        PUGI__FN xml_node xml_node::insert_copy_before(const xml_node& proto, const xml_node& node)
+        {
+                xml_node_type type_ = proto.type();
+                if (!impl::allow_insert_child(type(), type_)) return xml_node();
+                if (!node._root || node._root->parent != _root) return xml_node();
+
+                impl::xml_allocator& alloc = impl::get_allocator(_root);
+                if (!alloc.reserve()) return xml_node();
+
+                xml_node n(impl::allocate_node(alloc, type_));
+                if (!n) return xml_node();
+
+                impl::insert_node_before(n._root, node._root);
+                impl::node_copy_tree(n._root, proto._root);
+
+                return n;
+        }
+
+        PUGI__FN xml_node xml_node::append_move(const xml_node& moved)
+        {
+                if (!impl::allow_move(*this, moved)) return xml_node();
+
+                impl::xml_allocator& alloc = impl::get_allocator(_root);
+                if (!alloc.reserve()) return xml_node();
+
+                // disable document_buffer_order optimization since moving nodes around changes document order without changing buffer pointers
+                impl::get_document(_root).header |= impl::xml_memory_page_contents_shared_mask;
+
+                impl::remove_node(moved._root);
+                impl::append_node(moved._root, _root);
+
+                return moved;
+        }
+
+        PUGI__FN xml_node xml_node::prepend_move(const xml_node& moved)
+        {
+                if (!impl::allow_move(*this, moved)) return xml_node();
+
+                impl::xml_allocator& alloc = impl::get_allocator(_root);
+                if (!alloc.reserve()) return xml_node();
+
+                // disable document_buffer_order optimization since moving nodes around changes document order without changing buffer pointers
+                impl::get_document(_root).header |= impl::xml_memory_page_contents_shared_mask;
+
+                impl::remove_node(moved._root);
+                impl::prepend_node(moved._root, _root);
+
+                return moved;
+        }
+
+        PUGI__FN xml_node xml_node::insert_move_after(const xml_node& moved, const xml_node& node)
+        {
+                if (!impl::allow_move(*this, moved)) return xml_node();
+                if (!node._root || node._root->parent != _root) return xml_node();
+                if (moved._root == node._root) return xml_node();
+
+                impl::xml_allocator& alloc = impl::get_allocator(_root);
+                if (!alloc.reserve()) return xml_node();
+
+                // disable document_buffer_order optimization since moving nodes around changes document order without changing buffer pointers
+                impl::get_document(_root).header |= impl::xml_memory_page_contents_shared_mask;
+
+                impl::remove_node(moved._root);
+                impl::insert_node_after(moved._root, node._root);
+
+                return moved;
+        }
+
+        PUGI__FN xml_node xml_node::insert_move_before(const xml_node& moved, const xml_node& node)
+        {
+                if (!impl::allow_move(*this, moved)) return xml_node();
+                if (!node._root || node._root->parent != _root) return xml_node();
+                if (moved._root == node._root) return xml_node();
+
+                impl::xml_allocator& alloc = impl::get_allocator(_root);
+                if (!alloc.reserve()) return xml_node();
+
+                // disable document_buffer_order optimization since moving nodes around changes document order without changing buffer pointers
+                impl::get_document(_root).header |= impl::xml_memory_page_contents_shared_mask;
+
+                impl::remove_node(moved._root);
+                impl::insert_node_before(moved._root, node._root);
+
+                return moved;
+        }
+
+        PUGI__FN bool xml_node::remove_attribute(const char_t* name_)
+        {
+                return remove_attribute(attribute(name_));
+        }
+
+        PUGI__FN bool xml_node::remove_attribute(const xml_attribute& a)
+        {
+                if (!_root || !a._attr) return false;
+                if (!impl::is_attribute_of(a._attr, _root)) return false;
+
+                impl::xml_allocator& alloc = impl::get_allocator(_root);
+                if (!alloc.reserve()) return false;
+
+                impl::remove_attribute(a._attr, _root);
+                impl::destroy_attribute(a._attr, alloc);
+
+                return true;
+        }
+
+        PUGI__FN bool xml_node::remove_child(const char_t* name_)
+        {
+                return remove_child(child(name_));
+        }
+
+        PUGI__FN bool xml_node::remove_child(const xml_node& n)
+        {
+                if (!_root || !n._root || n._root->parent != _root) return false;
+
+                impl::xml_allocator& alloc = impl::get_allocator(_root);
+                if (!alloc.reserve()) return false;
+
+                impl::remove_node(n._root);
+                impl::destroy_node(n._root, alloc);
+
+                return true;
+        }
+
+        PUGI__FN xml_parse_result xml_node::append_buffer(const void* contents, size_t size, unsigned int options, xml_encoding encoding)
+        {
+                // append_buffer is only valid for elements/documents
+                if (!impl::allow_insert_child(type(), node_element)) return impl::make_parse_result(status_append_invalid_root);
+
+                // get document node
+                impl::xml_document_struct* doc = &impl::get_document(_root);
+
+                // disable document_buffer_order optimization since in a document with multiple buffers comparing buffer pointers does not make sense
+                doc->header |= impl::xml_memory_page_contents_shared_mask;
                 
-                if (node._root->prev_sibling_c->next_sibling)
-                        node._root->prev_sibling_c->next_sibling = n._root;
-                else
-                        _root->first_child = n._root;
-                
-                n._root->prev_sibling_c = node._root->prev_sibling_c;
-                n._root->next_sibling = node._root;
-                node._root->prev_sibling_c = n._root;
-
-                if (type_ == node_declaration) n.set_name(PUGIXML_TEXT("xml"));
-
-                return n;
-        }
-
-        PUGI__FN xml_node xml_node::insert_child_after(xml_node_type type_, const xml_node& node)
-        {
-                if (!impl::allow_insert_child(this->type(), type_)) return xml_node();
-                if (!node._root || node._root->parent != _root) return xml_node();
-        
-                xml_node n(impl::allocate_node(impl::get_allocator(_root), type_));
-                if (!n) return xml_node();
-
-                n._root->parent = _root;
-        
-                if (node._root->next_sibling)
-                        node._root->next_sibling->prev_sibling_c = n._root;
-                else
-                        _root->first_child->prev_sibling_c = n._root;
-                
-                n._root->next_sibling = node._root->next_sibling;
-                n._root->prev_sibling_c = node._root;
-                node._root->next_sibling = n._root;
-
-                if (type_ == node_declaration) n.set_name(PUGIXML_TEXT("xml"));
-
-                return n;
-        }
-
-        PUGI__FN xml_node xml_node::append_child(const char_t* name_)
-        {
-                xml_node result = append_child(node_element);
-
-                result.set_name(name_);
-
-                return result;
-        }
-
-        PUGI__FN xml_node xml_node::prepend_child(const char_t* name_)
-        {
-                xml_node result = prepend_child(node_element);
-
-                result.set_name(name_);
-
-                return result;
-        }
-
-        PUGI__FN xml_node xml_node::insert_child_after(const char_t* name_, const xml_node& node)
-        {
-                xml_node result = insert_child_after(node_element, node);
-
-                result.set_name(name_);
-
-                return result;
-        }
-
-        PUGI__FN xml_node xml_node::insert_child_before(const char_t* name_, const xml_node& node)
-        {
-                xml_node result = insert_child_before(node_element, node);
-
-                result.set_name(name_);
-
-                return result;
-        }
-
-        PUGI__FN xml_node xml_node::append_copy(const xml_node& proto)
-        {
-                xml_node result = append_child(proto.type());
-
-                if (result) impl::recursive_copy_skip(result, proto, result);
-
-                return result;
-        }
-
-        PUGI__FN xml_node xml_node::prepend_copy(const xml_node& proto)
-        {
-                xml_node result = prepend_child(proto.type());
-
-                if (result) impl::recursive_copy_skip(result, proto, result);
-
-                return result;
-        }
-
-        PUGI__FN xml_node xml_node::insert_copy_after(const xml_node& proto, const xml_node& node)
-        {
-                xml_node result = insert_child_after(proto.type(), node);
-
-                if (result) impl::recursive_copy_skip(result, proto, result);
-
-                return result;
-        }
-
-        PUGI__FN xml_node xml_node::insert_copy_before(const xml_node& proto, const xml_node& node)
-        {
-                xml_node result = insert_child_before(proto.type(), node);
-
-                if (result) impl::recursive_copy_skip(result, proto, result);
-
-                return result;
-        }
-
-        PUGI__FN bool xml_node::remove_attribute(const char_t* name_)
-        {
-                return remove_attribute(attribute(name_));
-        }
-
-        PUGI__FN bool xml_node::remove_attribute(const xml_attribute& a)
-        {
-                if (!_root || !a._attr) return false;
-
-                // check that attribute belongs to *this
-                xml_attribute_struct* attr = a._attr;
-
-                while (attr->prev_attribute_c->next_attribute) attr = attr->prev_attribute_c;
-
-                if (attr != _root->first_attribute) return false;
-
-                if (a._attr->next_attribute) a._attr->next_attribute->prev_attribute_c = a._attr->prev_attribute_c;
-                else if (_root->first_attribute) _root->first_attribute->prev_attribute_c = a._attr->prev_attribute_c;
-                
-                if (a._attr->prev_attribute_c->next_attribute) a._attr->prev_attribute_c->next_attribute = a._attr->next_attribute;
-                else _root->first_attribute = a._attr->next_attribute;
-
-                impl::destroy_attribute(a._attr, impl::get_allocator(_root));
-
-                return true;
-        }
-
-        PUGI__FN bool xml_node::remove_child(const char_t* name_)
-        {
-                return remove_child(child(name_));
-        }
-
-        PUGI__FN bool xml_node::remove_child(const xml_node& n)
-        {
-                if (!_root || !n._root || n._root->parent != _root) return false;
-
-                if (n._root->next_sibling) n._root->next_sibling->prev_sibling_c = n._root->prev_sibling_c;
-                else if (_root->first_child) _root->first_child->prev_sibling_c = n._root->prev_sibling_c;
-                
-                if (n._root->prev_sibling_c->next_sibling) n._root->prev_sibling_c->next_sibling = n._root->next_sibling;
-                else _root->first_child = n._root->next_sibling;
-                
-                impl::destroy_node(n._root, impl::get_allocator(_root));
-
-                return true;
+                // get extra buffer element (we'll store the document fragment buffer there so that we can deallocate it later)
+                impl::xml_memory_page* page = 0;
+                impl::xml_extra_buffer* extra = static_cast<impl::xml_extra_buffer*>(doc->allocate_memory(sizeof(impl::xml_extra_buffer), page));
+                (void)page;
+
+                if (!extra) return impl::make_parse_result(status_out_of_memory);
+
+                // add extra buffer to the list
+                extra->buffer = 0;
+                extra->next = doc->extra_buffers;
+                doc->extra_buffers = extra;
+
+                // name of the root has to be NULL before parsing - otherwise closing node mismatches will not be detected at the top level
+                impl::name_null_sentry sentry(_root);
+
+                return impl::load_buffer_impl(doc, _root, const_cast<void*>(contents), size, options, encoding, false, false, &extra->buffer);
         }
 
@@ -4508 +5955 @@
                         {
                                 for (xml_attribute_struct* a = i->first_attribute; a; a = a->next_attribute)
-                                        if (impl::strequal(attr_name, a->name) && impl::strequal(attr_value, a->value))
+                                        if (a->name && impl::strequal(attr_name, a->name) && impl::strequal(attr_value, a->value ? a->value + 0 : PUGIXML_TEXT("")))
                                                 return xml_node(i);
                         }
@@ -4521 +5968 @@
                 for (xml_node_struct* i = _root->first_child; i; i = i->next_sibling)
                         for (xml_attribute_struct* a = i->first_attribute; a; a = a->next_attribute)
-                                if (impl::strequal(attr_name, a->name) && impl::strequal(attr_value, a->value))
+                                if (a->name && impl::strequal(attr_name, a->name) && impl::strequal(attr_value, a->value ? a->value + 0 : PUGIXML_TEXT("")))
                                         return xml_node(i);
 
@@ -4530 +5977 @@
         PUGI__FN string_t xml_node::path(char_t delimiter) const
         {
-                xml_node cursor = *this; // Make a copy.
-                
-                string_t result = cursor.name();
-
-                while (cursor.parent())
-                {
-                        cursor = cursor.parent();
-                        
-                        string_t temp = cursor.name();
-                        temp += delimiter;
-                        temp += result;
-                        result.swap(temp);
-                }
+                if (!_root) return string_t();
+
+                size_t offset = 0;
+
+                for (xml_node_struct* i = _root; i; i = i->parent)
+                {
+                        offset += (i != _root);
+                        offset += i->name ? impl::strlength(i->name) : 0;
+                }
+
+                string_t result;
+                result.resize(offset);
+
+                for (xml_node_struct* j = _root; j; j = j->parent)
+                {
+                        if (j != _root)
+                                result[--offset] = delimiter;
+
+                        if (j->name && *j->name)
+                        {
+                                size_t length = impl::strlength(j->name);
+
+                                offset -= length;
+                                memcpy(&result[offset], j->name, length * sizeof(char_t));
+                        }
+                }
+
+                assert(offset == 0);
 
                 return result;
@@ -4659 +6121 @@
                 impl::xml_buffered_writer buffered_writer(writer, encoding);
 
-                impl::node_output(buffered_writer, *this, indent, flags, depth);
+                impl::node_output(buffered_writer, _root, indent, flags, depth);
+
+                buffered_writer.flush();
         }
 
@@ -4680 +6144 @@
         PUGI__FN ptrdiff_t xml_node::offset_debug() const
         {
-                xml_node_struct* r = root()._root;
-
-                if (!r) return -1;
-
-                const char_t* buffer = static_cast<impl::xml_document_struct*>(r)->buffer;
-
-                if (!buffer) return -1;
+                if (!_root) return -1;
+
+                impl::xml_document_struct& doc = impl::get_document(_root);
+
+                // we can determine the offset reliably only if there is exactly once parse buffer
+                if (!doc.buffer || doc.extra_buffers) return -1;
 
                 switch (type())
@@ -4696 +6159 @@
                 case node_declaration:
                 case node_pi:
-                        return (_root->header & impl::xml_memory_page_name_allocated_mask) ? -1 : _root->name - buffer;
+                        return _root->name && (_root->header & impl::xml_memory_page_name_allocated_or_shared_mask) == 0 ? _root->name - doc.buffer : -1;
 
                 case node_pcdata:
@@ -4702 +6165 @@
                 case node_comment:
                 case node_doctype:
-                        return (_root->header & impl::xml_memory_page_value_allocated_mask) ? -1 : _root->value - buffer;
+                        return _root->value && (_root->header & impl::xml_memory_page_value_allocated_or_shared_mask) == 0 ? _root->value - doc.buffer : -1;
 
                 default:
@@ -4771 +6234 @@
                 xml_node_struct* d = _data();
 
-                return (d && d->value) ? d->value : PUGIXML_TEXT("");
+                return (d && d->value) ? d->value + 0 : PUGIXML_TEXT("");
         }
 
@@ -4778 +6241 @@
                 xml_node_struct* d = _data();
 
-                return (d && d->value) ? d->value : def;
+                return (d && d->value) ? d->value + 0 : def;
         }
 
@@ -4785 +6248 @@
                 xml_node_struct* d = _data();
 
-                return impl::get_value_int(d ? d->value : 0, def);
+                return (d && d->value) ? impl::get_value_int(d->value) : def;
         }
 
@@ -4792 +6255 @@
                 xml_node_struct* d = _data();
 
-                return impl::get_value_uint(d ? d->value : 0, def);
+                return (d && d->value) ? impl::get_value_uint(d->value) : def;
         }
 
@@ -4799 +6262 @@
                 xml_node_struct* d = _data();
 
-                return impl::get_value_double(d ? d->value : 0, def);
+                return (d && d->value) ? impl::get_value_double(d->value) : def;
         }
 
@@ -4806 +6269 @@
                 xml_node_struct* d = _data();
 
-                return impl::get_value_float(d ? d->value : 0, def);
+                return (d && d->value) ? impl::get_value_float(d->value) : def;
         }
 
@@ -4813 +6276 @@
                 xml_node_struct* d = _data();
 
-                return impl::get_value_bool(d ? d->value : 0, def);
-        }
+                return (d && d->value) ? impl::get_value_bool(d->value) : def;
+        }
+
+#ifdef PUGIXML_HAS_LONG_LONG
+        PUGI__FN long long xml_text::as_llong(long long def) const
+        {
+                xml_node_struct* d = _data();
+
+                return (d && d->value) ? impl::get_value_llong(d->value) : def;
+        }
+
+        PUGI__FN unsigned long long xml_text::as_ullong(unsigned long long def) const
+        {
+                xml_node_struct* d = _data();
+
+                return (d && d->value) ? impl::get_value_ullong(d->value) : def;
+        }
+#endif
 
         PUGI__FN bool xml_text::set(const char_t* rhs)
@@ -4820 +6299 @@
                 xml_node_struct* dn = _data_new();
 
-                return dn ? impl::strcpy_insitu(dn->value, dn->header, impl::xml_memory_page_value_allocated_mask, rhs) : false;
+                return dn ? impl::strcpy_insitu(dn->value, dn->header, impl::xml_memory_page_value_allocated_mask, rhs, impl::strlength(rhs)) : false;
         }
 
@@ -4837 +6316 @@
         }
 
+        PUGI__FN bool xml_text::set(float rhs)
+        {
+                xml_node_struct* dn = _data_new();
+
+                return dn ? impl::set_value_convert(dn->value, dn->header, impl::xml_memory_page_value_allocated_mask, rhs) : false;
+        }
+
         PUGI__FN bool xml_text::set(double rhs)
         {
@@ -4850 +6336 @@
                 return dn ? impl::set_value_convert(dn->value, dn->header, impl::xml_memory_page_value_allocated_mask, rhs) : false;
         }
+
+#ifdef PUGIXML_HAS_LONG_LONG
+        PUGI__FN bool xml_text::set(long long rhs)
+        {
+                xml_node_struct* dn = _data_new();
+
+                return dn ? impl::set_value_convert(dn->value, dn->header, impl::xml_memory_page_value_allocated_mask, rhs) : false;
+        }
+
+        PUGI__FN bool xml_text::set(unsigned long long rhs)
+        {
+                xml_node_struct* dn = _data_new();
+
+                return dn ? impl::set_value_convert(dn->value, dn->header, impl::xml_memory_page_value_allocated_mask, rhs) : false;
+        }
+#endif
 
         PUGI__FN xml_text& xml_text::operator=(const char_t* rhs)
@@ -4875 +6377 @@
         }
 
-        PUGI__FN xml_text& xml_text::operator=(bool rhs)
+        PUGI__FN xml_text& xml_text::operator=(float rhs)
         {
                 set(rhs);
                 return *this;
         }
+
+        PUGI__FN xml_text& xml_text::operator=(bool rhs)
+        {
+                set(rhs);
+                return *this;
+        }
+
+#ifdef PUGIXML_HAS_LONG_LONG
+        PUGI__FN xml_text& xml_text::operator=(long long rhs)
+        {
+                set(rhs);
+                return *this;
+        }
+
+        PUGI__FN xml_text& xml_text::operator=(unsigned long long rhs)
+        {
+                set(rhs);
+                return *this;
+        }
+#endif
 
         PUGI__FN xml_node xml_text::data() const
@@ -5024 +6546 @@
         }
 
-        PUGI__FN xml_named_node_iterator::xml_named_node_iterator(const xml_node& node, const char_t* name): _node(node), _name(name)
+        PUGI__FN xml_named_node_iterator::xml_named_node_iterator(const xml_node& node, const char_t* name): _wrap(node), _parent(node.parent()), _name(name)
+        {
+        }
+
+        PUGI__FN xml_named_node_iterator::xml_named_node_iterator(xml_node_struct* ref, xml_node_struct* parent, const char_t* name): _wrap(ref), _parent(parent), _name(name)
         {
         }
@@ -5030 +6556 @@
         PUGI__FN bool xml_named_node_iterator::operator==(const xml_named_node_iterator& rhs) const
         {
-                return _node == rhs._node;
+                return _wrap._root == rhs._wrap._root && _parent._root == rhs._parent._root;
         }
 
         PUGI__FN bool xml_named_node_iterator::operator!=(const xml_named_node_iterator& rhs) const
         {
-                return _node != rhs._node;
+                return _wrap._root != rhs._wrap._root || _parent._root != rhs._parent._root;
         }
 
         PUGI__FN xml_node& xml_named_node_iterator::operator*() const
         {
-                assert(_node._root);
-                return _node;
+                assert(_wrap._root);
+                return _wrap;
         }
 
         PUGI__FN xml_node* xml_named_node_iterator::operator->() const
         {
-                assert(_node._root);
-                return const_cast<xml_node*>(&_node); // BCC32 workaround
+                assert(_wrap._root);
+                return const_cast<xml_node*>(&_wrap); // BCC32 workaround
         }
 
         PUGI__FN const xml_named_node_iterator& xml_named_node_iterator::operator++()
         {
-                assert(_node._root);
-                _node = _node.next_sibling(_name);
+                assert(_wrap._root);
+                _wrap = _wrap.next_sibling(_name);
                 return *this;
         }
@@ -5061 +6587 @@
                 xml_named_node_iterator temp = *this;
                 ++*this;
+                return temp;
+        }
+
+        PUGI__FN const xml_named_node_iterator& xml_named_node_iterator::operator--()
+        {
+                if (_wrap._root)
+                        _wrap = _wrap.previous_sibling(_name);
+                else
+                {
+                        _wrap = _parent.last_child();
+
+                        if (!impl::strequal(_wrap.name(), _name))
+                                _wrap = _wrap.previous_sibling(_name);
+                }
+
+                return *this;
+        }
+
+        PUGI__FN xml_named_node_iterator xml_named_node_iterator::operator--(int)
+        {
+                xml_named_node_iterator temp = *this;
+                --*this;
                 return temp;
         }
@@ -5096 +6644 @@
                 case status_end_element_mismatch: return "Start-end tags mismatch";
 
+                case status_append_invalid_root: return "Unable to append nodes: root is not an element or document";
+
+                case status_no_document_element: return "No document element found";
+
                 default: return "Unknown error";
                 }
@@ -5126 +6678 @@
         PUGI__FN void xml_document::create()
         {
+                assert(!_root);
+
+        #ifdef PUGIXML_COMPACT
+                const size_t page_offset = sizeof(uint32_t);
+        #else
+                const size_t page_offset = 0;
+        #endif
+
                 // initialize sentinel page
-                PUGI__STATIC_ASSERT(offsetof(impl::xml_memory_page, data) + sizeof(impl::xml_document_struct) + impl::xml_memory_page_alignment <= sizeof(_memory));
+                PUGI__STATIC_ASSERT(sizeof(impl::xml_memory_page) + sizeof(impl::xml_document_struct) + impl::xml_memory_page_alignment - sizeof(void*) + page_offset <= sizeof(_memory));
 
                 // align upwards to page boundary
@@ -5134 +6694 @@
                 // prepare page structure
                 impl::xml_memory_page* page = impl::xml_memory_page::construct(page_memory);
+                assert(page);
 
                 page->busy_size = impl::xml_memory_page_size;
 
+                // setup first page marker
+        #ifdef PUGIXML_COMPACT
+                // round-trip through void* to avoid 'cast increases required alignment of target type' warning
+                page->compact_page_marker = reinterpret_cast<uint32_t*>(static_cast<void*>(reinterpret_cast<char*>(page) + sizeof(impl::xml_memory_page)));
+                *page->compact_page_marker = sizeof(impl::xml_memory_page);
+        #endif
+
                 // allocate new root
-                _root = new (page->data) impl::xml_document_struct(page);
+                _root = new (reinterpret_cast<char*>(page) + sizeof(impl::xml_memory_page) + page_offset) impl::xml_document_struct(page);
                 _root->prev_sibling_c = _root;
 
                 // setup sentinel page
                 page->allocator = static_cast<impl::xml_document_struct*>(_root);
+
+                // verify the document allocation
+                assert(reinterpret_cast<char*>(_root) + sizeof(impl::xml_document_struct) <= _memory + sizeof(_memory));
         }
 
         PUGI__FN void xml_document::destroy()
         {
+                assert(_root);
+
                 // destroy static storage
                 if (_buffer)
@@ -5154 +6727 @@
                 }
 
+                // destroy extra buffers (note: no need to destroy linked list nodes, they're allocated using document allocator)
+                for (impl::xml_extra_buffer* extra = static_cast<impl::xml_document_struct*>(_root)->extra_buffers; extra; extra = extra->next)
+                {
+                        if (extra->buffer) impl::xml_memory::deallocate(extra->buffer);
+                }
+
                 // destroy dynamic storage, leave sentinel page (it's in static memory)
-                if (_root)
-                {
-                        impl::xml_memory_page* root_page = reinterpret_cast<impl::xml_memory_page*>(_root->header & impl::xml_memory_page_pointer_mask);
-                        assert(root_page && !root_page->prev && !root_page->memory);
-
-                        // destroy all pages
-                        for (impl::xml_memory_page* page = root_page->next; page; )
-                        {
-                                impl::xml_memory_page* next = page->next;
-
-                                impl::xml_allocator::deallocate_page(page);
-
-                                page = next;
-                        }
-
-                        // cleanup root page
-                        root_page->allocator = 0;
-                        root_page->next = 0;
-                        root_page->busy_size = root_page->freed_size = 0;
-
-                        _root = 0;
-                }
+                impl::xml_memory_page* root_page = PUGI__GETPAGE(_root);
+                assert(root_page && !root_page->prev);
+                assert(reinterpret_cast<char*>(root_page) >= _memory && reinterpret_cast<char*>(root_page) < _memory + sizeof(_memory));
+
+                for (impl::xml_memory_page* page = root_page->next; page; )
+                {
+                        impl::xml_memory_page* next = page->next;
+
+                        impl::xml_allocator::deallocate_page(page);
+
+                        page = next;
+                }
+
+        #ifdef PUGIXML_COMPACT
+                // destroy hash table
+                static_cast<impl::xml_document_struct*>(_root)->hash.clear();
+        #endif
+
+                _root = 0;
         }
 
@@ -5184 +6760 @@
                 reset();
 
-                return impl::load_stream_impl(*this, stream, options, encoding);
+                return impl::load_stream_impl(static_cast<impl::xml_document_struct*>(_root), stream, options, encoding, &_buffer);
         }
 
@@ -5191 +6767 @@
                 reset();
 
-                return impl::load_stream_impl(*this, stream, options, encoding_wchar);
+                return impl::load_stream_impl(static_cast<impl::xml_document_struct*>(_root), stream, options, encoding_wchar, &_buffer);
         }
 #endif
 
-        PUGI__FN xml_parse_result xml_document::load(const char_t* contents, unsigned int options)
+        PUGI__FN xml_parse_result xml_document::load_string(const char_t* contents, unsigned int options)
         {
                 // Force native encoding (skip autodetection)
@@ -5207 +6783 @@
         }
 
+        PUGI__FN xml_parse_result xml_document::load(const char_t* contents, unsigned int options)
+        {
+                return load_string(contents, options);
+        }
+
         PUGI__FN xml_parse_result xml_document::load_file(const char* path_, unsigned int options, xml_encoding encoding)
         {
                 reset();
 
-                FILE* file = fopen(path_, "rb");
-
-                return impl::load_file_impl(*this, file, options, encoding);
+                using impl::auto_deleter; // MSVC7 workaround
+                auto_deleter<FILE, int(*)(FILE*)> file(fopen(path_, "rb"), fclose);
+
+                return impl::load_file_impl(static_cast<impl::xml_document_struct*>(_root), file.data, options, encoding, &_buffer);
         }
 
@@ -5220 +6802 @@
                 reset();
 
-                FILE* file = impl::open_file_wide(path_, L"rb");
-
-                return impl::load_file_impl(*this, file, options, encoding);
-        }
-
-        PUGI__FN xml_parse_result xml_document::load_buffer_impl(void* contents, size_t size, unsigned int options, xml_encoding encoding, bool is_mutable, bool own)
+                using impl::auto_deleter; // MSVC7 workaround
+                auto_deleter<FILE, int(*)(FILE*)> file(impl::open_file_wide(path_, L"rb"), fclose);
+
+                return impl::load_file_impl(static_cast<impl::xml_document_struct*>(_root), file.data, options, encoding, &_buffer);
+        }
+
+        PUGI__FN xml_parse_result xml_document::load_buffer(const void* contents, size_t size, unsigned int options, xml_encoding encoding)
         {
                 reset();
 
-                // check input buffer
-                assert(contents || size == 0);
-
-                // get actual encoding
-                xml_encoding buffer_encoding = impl::get_buffer_encoding(encoding, contents, size);
-
-                // get private buffer
-                char_t* buffer = 0;
-                size_t length = 0;
-
-                if (!impl::convert_buffer(buffer, length, buffer_encoding, contents, size, is_mutable)) return impl::make_parse_result(status_out_of_memory);
-                
-                // delete original buffer if we performed a conversion
-                if (own && buffer != contents && contents) impl::xml_memory::deallocate(contents);
-
-                // parse
-                xml_parse_result res = impl::xml_parser::parse(buffer, length, _root, options);
-
-                // remember encoding
-                res.encoding = buffer_encoding;
-
-                // grab onto buffer if it's our buffer, user is responsible for deallocating contens himself
-                if (own || buffer != contents) _buffer = buffer;
-
-                return res;
-        }
-
-        PUGI__FN xml_parse_result xml_document::load_buffer(const void* contents, size_t size, unsigned int options, xml_encoding encoding)
-        {
-                return load_buffer_impl(const_cast<void*>(contents), size, options, encoding, false, false);
+                return impl::load_buffer_impl(static_cast<impl::xml_document_struct*>(_root), _root, const_cast<void*>(contents), size, options, encoding, false, false, &_buffer);
         }
 
         PUGI__FN xml_parse_result xml_document::load_buffer_inplace(void* contents, size_t size, unsigned int options, xml_encoding encoding)
         {
-                return load_buffer_impl(contents, size, options, encoding, true, false);
-        }
-                
+                reset();
+
+                return impl::load_buffer_impl(static_cast<impl::xml_document_struct*>(_root), _root, contents, size, options, encoding, true, false, &_buffer);
+        }
+
         PUGI__FN xml_parse_result xml_document::load_buffer_inplace_own(void* contents, size_t size, unsigned int options, xml_encoding encoding)
         {
-                return load_buffer_impl(contents, size, options, encoding, true, true);
+                reset();
+
+                return impl::load_buffer_impl(static_cast<impl::xml_document_struct*>(_root), _root, contents, size, options, encoding, true, true, &_buffer);
         }
 
@@ -5286 +6844 @@
                 }
 
-                if (!(flags & format_no_declaration) && !impl::has_declaration(*this))
-                {
-                        buffered_writer.write(PUGIXML_TEXT("<?xml version=\"1.0\""));
-                        if (encoding == encoding_latin1) buffered_writer.write(PUGIXML_TEXT(" encoding=\"ISO-8859-1\""));
+                if (!(flags & format_no_declaration) && !impl::has_declaration(_root))
+                {
+                        buffered_writer.write_string(PUGIXML_TEXT("<?xml version=\"1.0\""));
+                        if (encoding == encoding_latin1) buffered_writer.write_string(PUGIXML_TEXT(" encoding=\"ISO-8859-1\""));
                         buffered_writer.write('?', '>');
                         if (!(flags & format_raw)) buffered_writer.write('\n');
                 }
 
-                impl::node_output(buffered_writer, *this, indent, flags, 0);
+                impl::node_output(buffered_writer, _root, indent, flags, 0);
+
+                buffered_writer.flush();
         }
 
@@ -5315 +6875 @@
         PUGI__FN bool xml_document::save_file(const char* path_, const char_t* indent, unsigned int flags, xml_encoding encoding) const
         {
-                FILE* file = fopen(path_, (flags & format_save_file_text) ? "w" : "wb");
-                return impl::save_file_impl(*this, file, indent, flags, encoding);
+                using impl::auto_deleter; // MSVC7 workaround
+                auto_deleter<FILE, int(*)(FILE*)> file(fopen(path_, (flags & format_save_file_text) ? "w" : "wb"), fclose);
+
+                return impl::save_file_impl(*this, file.data, indent, flags, encoding);
         }
 
         PUGI__FN bool xml_document::save_file(const wchar_t* path_, const char_t* indent, unsigned int flags, xml_encoding encoding) const
         {
-                FILE* file = impl::open_file_wide(path_, (flags & format_save_file_text) ? L"w" : L"wb");
-                return impl::save_file_impl(*this, file, indent, flags, encoding);
+                using impl::auto_deleter; // MSVC7 workaround
+                auto_deleter<FILE, int(*)(FILE*)> file(impl::open_file_wide(path_, (flags & format_save_file_text) ? L"w" : L"wb"), fclose);
+
+                return impl::save_file_impl(*this, file.data, indent, flags, encoding);
         }
 
         PUGI__FN xml_node xml_document::document_element() const
         {
+                assert(_root);
+
                 for (xml_node_struct* i = _root->first_child; i; i = i->next_sibling)
-                        if ((i->header & impl::xml_memory_page_type_mask) + 1 == node_element)
+                        if (PUGI__NODETYPE(i) == node_element)
                                 return xml_node(i);
 
@@ -5339 +6905 @@
                 assert(str);
 
-                return impl::as_utf8_impl(str, wcslen(str));
+                return impl::as_utf8_impl(str, impl::strlength_wide(str));
         }
 
@@ -5391 +6957 @@
         }
 
-        PUGI__FN std::forward_iterator_tag _Iter_cat(const pugi::xml_named_node_iterator&)
-        {
-                return std::forward_iterator_tag();
+        PUGI__FN std::bidirectional_iterator_tag _Iter_cat(const pugi::xml_named_node_iterator&)
+        {
+                return std::bidirectional_iterator_tag();
         }
 }
@@ -5412 +6978 @@
         }
 
-        PUGI__FN std::forward_iterator_tag __iterator_category(const pugi::xml_named_node_iterator&)
-        {
-                return std::forward_iterator_tag();
+        PUGI__FN std::bidirectional_iterator_tag __iterator_category(const pugi::xml_named_node_iterator&)
+        {
+                return std::bidirectional_iterator_tag();
         }
 }
@@ -5420 +6986 @@
 
 #ifndef PUGIXML_NO_XPATH
-
 // STL replacements
 PUGI__NS_BEGIN
@@ -5475 +7040 @@
         template <typename I> void reverse(I begin, I end)
         {
-                while (begin + 1 < end) swap(*begin++, *--end);
+                while (end - begin > 1) swap(*begin++, *--end);
         }
 
@@ -5481 +7046 @@
         {
                 // fast skip head
-                while (begin + 1 < end && *begin != *(begin + 1)) begin++;
+                while (end - begin > 1 && *begin != *(begin + 1)) begin++;
 
                 if (begin == end) return begin;
@@ -5650 +7215 @@
 // Allocator used for AST and evaluation stacks
 PUGI__NS_BEGIN
+        static const size_t xpath_memory_page_size =
+        #ifdef PUGIXML_MEMORY_XPATH_PAGE_SIZE
+                PUGIXML_MEMORY_XPATH_PAGE_SIZE
+        #else
+                4096
+        #endif
+                ;
+
+        static const uintptr_t xpath_memory_block_alignment = sizeof(double) > sizeof(void*) ? sizeof(double) : sizeof(void*);
+
         struct xpath_memory_block
         {       
                 xpath_memory_block* next;
-
-                char data[
-        #ifdef PUGIXML_MEMORY_XPATH_PAGE_SIZE
-                        PUGIXML_MEMORY_XPATH_PAGE_SIZE
-        #else
-                        4096
-        #endif
-                ];
+                size_t capacity;
+
+                union
+                {
+                        char data[xpath_memory_page_size];
+                        double alignment;
+                };
         };
                 
@@ -5682 +7256 @@
                 void* allocate_nothrow(size_t size)
                 {
-                        const size_t block_capacity = sizeof(_root->data);
-
-                        // align size so that we're able to store pointers in subsequent blocks
-                        size = (size + sizeof(void*) - 1) & ~(sizeof(void*) - 1);
-
-                        if (_root_size + size <= block_capacity)
-                        {
-                                void* buf = _root->data + _root_size;
+                        // round size up to block alignment boundary
+                        size = (size + xpath_memory_block_alignment - 1) & ~(xpath_memory_block_alignment - 1);
+
+                        if (_root_size + size <= _root->capacity)
+                        {
+                                void* buf = &_root->data[0] + _root_size;
                                 _root_size += size;
                                 return buf;
@@ -5695 +7267 @@
                         else
                         {
-                                size_t block_data_size = (size > block_capacity) ? size : block_capacity;
-                                size_t block_size = block_data_size + offsetof(xpath_memory_block, data);
+                                // make sure we have at least 1/4th of the page free after allocation to satisfy subsequent allocation requests
+                                size_t block_capacity_base = sizeof(_root->data);
+                                size_t block_capacity_req = size + block_capacity_base / 4;
+                                size_t block_capacity = (block_capacity_base > block_capacity_req) ? block_capacity_base : block_capacity_req;
+
+                                size_t block_size = block_capacity + offsetof(xpath_memory_block, data);
 
                                 xpath_memory_block* block = static_cast<xpath_memory_block*>(xml_memory::allocate(block_size));
@@ -5702 +7278 @@
                                 
                                 block->next = _root;
+                                block->capacity = block_capacity;
                                 
                                 _root = block;
@@ -5729 +7306 @@
                 void* reallocate(void* ptr, size_t old_size, size_t new_size)
                 {
-                        // align size so that we're able to store pointers in subsequent blocks
-                        old_size = (old_size + sizeof(void*) - 1) & ~(sizeof(void*) - 1);
-                        new_size = (new_size + sizeof(void*) - 1) & ~(sizeof(void*) - 1);
+                        // round size up to block alignment boundary
+                        old_size = (old_size + xpath_memory_block_alignment - 1) & ~(xpath_memory_block_alignment - 1);
+                        new_size = (new_size + xpath_memory_block_alignment - 1) & ~(xpath_memory_block_alignment - 1);
 
                         // we can only reallocate the last object
-                        assert(ptr == 0 || static_cast<char*>(ptr) + old_size == _root->data + _root_size);
+                        assert(ptr == 0 || static_cast<char*>(ptr) + old_size == &_root->data[0] + _root_size);
 
                         // adjust root size so that we have not allocated the object at all
@@ -5749 +7326 @@
                         {
                                 // copy old data
-                                assert(new_size > old_size);
+                                assert(new_size >= old_size);
                                 memcpy(result, ptr, old_size);
 
@@ -5842 +7419 @@
                 {
                         blocks[0].next = blocks[1].next = 0;
+                        blocks[0].capacity = blocks[1].capacity = sizeof(blocks[0].data);
 
                         stack.result = &result;
@@ -5865 +7443 @@
                 const char_t* _buffer;
                 bool _uses_heap;
+                size_t _length_heap;
 
                 static char_t* duplicate_string(const char_t* string, size_t length, xpath_allocator* alloc)
@@ -5877 +7456 @@
                 }
 
-                static char_t* duplicate_string(const char_t* string, xpath_allocator* alloc)
-                {
-                        return duplicate_string(string, strlength(string), alloc);
+                xpath_string(const char_t* buffer, bool uses_heap_, size_t length_heap): _buffer(buffer), _uses_heap(uses_heap_), _length_heap(length_heap)
+                {
                 }
 
         public:
-                xpath_string(): _buffer(PUGIXML_TEXT("")), _uses_heap(false)
-                {
-                }
-
-                explicit xpath_string(const char_t* str, xpath_allocator* alloc)
-                {
-                        bool empty_ = (*str == 0);
-
-                        _buffer = empty_ ? PUGIXML_TEXT("") : duplicate_string(str, alloc);
-                        _uses_heap = !empty_;
-                }
-
-                explicit xpath_string(const char_t* str, bool use_heap): _buffer(str), _uses_heap(use_heap)
-                {
-                }
-
-                xpath_string(const char_t* begin, const char_t* end, xpath_allocator* alloc)
+                static xpath_string from_const(const char_t* str)
+                {
+                        return xpath_string(str, false, 0);
+                }
+
+                static xpath_string from_heap_preallocated(const char_t* begin, const char_t* end)
+                {
+                        assert(begin <= end && *end == 0);
+
+                        return xpath_string(begin, true, static_cast<size_t>(end - begin));
+                }
+
+                static xpath_string from_heap(const char_t* begin, const char_t* end, xpath_allocator* alloc)
                 {
                         assert(begin <= end);
 
-                        bool empty_ = (begin == end);
-
-                        _buffer = empty_ ? PUGIXML_TEXT("") : duplicate_string(begin, static_cast<size_t>(end - begin), alloc);
-                        _uses_heap = !empty_;
+                        size_t length = static_cast<size_t>(end - begin);
+
+                        return length == 0 ? xpath_string() : xpath_string(duplicate_string(begin, length, alloc), true, length);
+                }
+
+                xpath_string(): _buffer(PUGIXML_TEXT("")), _uses_heap(false), _length_heap(0)
+                {
                 }
 
@@ -5922 +7499 @@
                         {
                                 // need to make heap copy
-                                size_t target_length = strlength(_buffer);
-                                size_t source_length = strlength(o._buffer);
+                                size_t target_length = length();
+                                size_t source_length = o.length();
                                 size_t result_length = target_length + source_length;
 
@@ -5940 +7517 @@
                                 _buffer = result;
                                 _uses_heap = true;
+                                _length_heap = result_length;
                         }
                 }
@@ -5950 +7528 @@
                 size_t length() const
                 {
-                        return strlength(_buffer);
+                        return _uses_heap ? _length_heap : strlength(_buffer);
                 }
                 
@@ -5958 +7536 @@
                         if (!_uses_heap)
                         {
-                                _buffer = duplicate_string(_buffer, alloc);
+                                size_t length_ = strlength(_buffer);
+
+                                _buffer = duplicate_string(_buffer, length_, alloc);
                                 _uses_heap = true;
+                                _length_heap = length_;
                         }
 
@@ -5985 +7566 @@
                 }
         };
-
-        PUGI__FN xpath_string xpath_string_const(const char_t* str)
-        {
-                return xpath_string(str, false);
-        }
 PUGI__NS_END
 
@@ -6032 +7608 @@
         {
                 if (na.attribute())
-                        return xpath_string_const(na.attribute().value());
+                        return xpath_string::from_const(na.attribute().value());
                 else
                 {
-                        const xml_node& n = na.node();
+                        xml_node n = na.node();
 
                         switch (n.type())
@@ -6043 +7619 @@
                         case node_comment:
                         case node_pi:
-                                return xpath_string_const(n.value());
+                                return xpath_string::from_const(n.value());
                         
                         case node_document:
@@ -6055 +7631 @@
                                 {
                                         if (cur.type() == node_pcdata || cur.type() == node_cdata)
-                                                result.append(xpath_string_const(cur.value()), alloc);
+                                                result.append(xpath_string::from_const(cur.value()), alloc);
 
                                         if (cur.first_child())
@@ -6079 +7655 @@
         }
         
-        PUGI__FN unsigned int node_height(xml_node n)
-        {
-                unsigned int result = 0;
-                
-                while (n)
-                {
-                        ++result;
-                        n = n.parent();
-                }
-                
-                return result;
+        PUGI__FN bool node_is_before_sibling(xml_node_struct* ln, xml_node_struct* rn)
+        {
+                assert(ln->parent == rn->parent);
+
+                // there is no common ancestor (the shared parent is null), nodes are from different documents
+                if (!ln->parent) return ln < rn;
+
+                // determine sibling order
+                xml_node_struct* ls = ln;
+                xml_node_struct* rs = rn;
+
+                while (ls && rs)
+                {
+                        if (ls == rn) return true;
+                        if (rs == ln) return false;
+
+                        ls = ls->next_sibling;
+                        rs = rs->next_sibling;
+                }
+
+                // if rn sibling chain ended ln must be before rn
+                return !rs;
         }
         
-        PUGI__FN bool node_is_before(xml_node ln, unsigned int lh, xml_node rn, unsigned int rh)
-        {
-                // normalize heights
-                for (unsigned int i = rh; i < lh; i++) ln = ln.parent();
-                for (unsigned int j = lh; j < rh; j++) rn = rn.parent();
-                
+        PUGI__FN bool node_is_before(xml_node_struct* ln, xml_node_struct* rn)
+        {
+                // find common ancestor at the same depth, if any
+                xml_node_struct* lp = ln;
+                xml_node_struct* rp = rn;
+
+                while (lp && rp && lp->parent != rp->parent)
+                {
+                        lp = lp->parent;
+                        rp = rp->parent;
+                }
+
+                // parents are the same!
+                if (lp && rp) return node_is_before_sibling(lp, rp);
+
+                // nodes are at different depths, need to normalize heights
+                bool left_higher = !lp;
+
+                while (lp)
+                {
+                        lp = lp->parent;
+                        ln = ln->parent;
+                }
+
+                while (rp)
+                {
+                        rp = rp->parent;
+                        rn = rn->parent;
+                }
+
                 // one node is the ancestor of the other
-                if (ln == rn) return lh < rh;
-                
-                // find common ancestor
-                while (ln.parent() != rn.parent())
-                {
-                        ln = ln.parent();
-                        rn = rn.parent();
-                }
-
-                // there is no common ancestor (the shared parent is null), nodes are from different documents
-                if (!ln.parent()) return ln < rn;
-
-                // determine sibling order
-                for (; ln; ln = ln.next_sibling())
-                        if (ln == rn)
-                                return true;
-                                
-                return false;
-        }
-
-        PUGI__FN bool node_is_ancestor(xml_node parent, xml_node node)
-        {
-                while (node && node != parent) node = node.parent();
+                if (ln == rn) return left_higher;
+
+                // find common ancestor... again
+                while (ln->parent != rn->parent)
+                {
+                        ln = ln->parent;
+                        rn = rn->parent;
+                }
+
+                return node_is_before_sibling(ln, rn);
+        }
+
+        PUGI__FN bool node_is_ancestor(xml_node_struct* parent, xml_node_struct* node)
+        {
+                while (node && node != parent) node = node->parent;
 
                 return parent && node == parent;
         }
 
-        PUGI__FN const void* document_order(const xpath_node& xnode)
+        PUGI__FN const void* document_buffer_order(const xpath_node& xnode)
         {
                 xml_node_struct* node = xnode.node().internal_object();
@@ -6132 +7735 @@
                 if (node)
                 {
-                        if (node->name && (node->header & xml_memory_page_name_allocated_mask) == 0) return node->name;
-                        if (node->value && (node->header & xml_memory_page_value_allocated_mask) == 0) return node->value;
+                        if ((get_document(node).header & xml_memory_page_contents_shared_mask) == 0)
+                        {
+                                if (node->name && (node->header & impl::xml_memory_page_name_allocated_or_shared_mask) == 0) return node->name;
+                                if (node->value && (node->header & impl::xml_memory_page_value_allocated_or_shared_mask) == 0) return node->value;
+                        }
+
                         return 0;
                 }
@@ -6141 +7748 @@
                 if (attr)
                 {
-                        if ((attr->header & xml_memory_page_name_allocated_mask) == 0) return attr->name;
-                        if ((attr->header & xml_memory_page_value_allocated_mask) == 0) return attr->value;
+                        if ((get_document(attr).header & xml_memory_page_contents_shared_mask) == 0)
+                        {
+                                if ((attr->header & impl::xml_memory_page_name_allocated_or_shared_mask) == 0) return attr->name;
+                                if ((attr->header & impl::xml_memory_page_value_allocated_or_shared_mask) == 0) return attr->value;
+                        }
+
                         return 0;
                 }
@@ -6154 +7765 @@
                 {
                         // optimized document order based check
-                        const void* lo = document_order(lhs);
-                        const void* ro = document_order(rhs);
+                        const void* lo = document_buffer_order(lhs);
+                        const void* ro = document_buffer_order(rhs);
 
                         if (lo && ro) return lo < ro;
@@ -6196 +7807 @@
 
                         if (ln == rn) return false;
+
+                        if (!ln || !rn) return ln < rn;
                         
-                        unsigned int lh = node_height(ln);
-                        unsigned int rh = node_height(rn);
-                        
-                        return node_is_before(ln, lh, rn, rh);
+                        return node_is_before(ln.internal_object(), rn.internal_object());
                 }
         };
@@ -6334 +7944 @@
                 // try special number conversion
                 const char_t* special = convert_number_to_string_special(value);
-                if (special) return xpath_string_const(special);
+                if (special) return xpath_string::from_const(special);
 
                 // get mantissa + exponent form
-                char mantissa_buffer[64];
+                char mantissa_buffer[32];
 
                 char* mantissa;
@@ -6343 +7953 @@
                 convert_number_to_mantissa_exponent(value, mantissa_buffer, sizeof(mantissa_buffer), &mantissa, &exponent);
 
+                // allocate a buffer of suitable length for the number
+                size_t result_size = strlen(mantissa_buffer) + (exponent > 0 ? exponent : -exponent) + 4;
+                char_t* result = static_cast<char_t*>(alloc->allocate(sizeof(char_t) * result_size));
+                assert(result);
+
                 // make the number!
-                char_t result[512];
                 char_t* s = result;
 
@@ -6359 +7973 @@
                         while (exponent > 0)
                         {
-                                assert(*mantissa == 0 || static_cast<unsigned int>(*mantissa - '0') <= 9);
+                                assert(*mantissa == 0 || static_cast<unsigned int>(static_cast<unsigned int>(*mantissa) - '0') <= 9);
                                 *s++ = *mantissa ? *mantissa++ : '0';
                                 exponent--;
@@ -6387 +8001 @@
 
                 // zero-terminate
-                assert(s < result + sizeof(result) / sizeof(result[0]));
+                assert(s < result + result_size);
                 *s = 0;
 
-                return xpath_string(result, alloc);
+                return xpath_string::from_heap_preallocated(result, s);
         }
         
@@ -6432 +8046 @@
                 return wcstod(string, 0);
         #else
-                return atof(string);
+                return strtod(string, 0);
         #endif
         }
 
-        PUGI__FN bool convert_string_to_number(const char_t* begin, const char_t* end, double* out_result)
-        {
-                char_t buffer[32];
-
+        PUGI__FN bool convert_string_to_number_scratch(char_t (&buffer)[32], const char_t* begin, const char_t* end, double* out_result)
+        {
                 size_t length = static_cast<size_t>(end - begin);
                 char_t* scratch = buffer;
@@ -6500 +8112 @@
                 }
 
-                bool operator()(const xml_attribute& a) const
+                bool operator()(xml_attribute a) const
                 {
                         const char_t* name = a.name();
@@ -6510 +8122 @@
         };
 
-        PUGI__FN const char_t* namespace_uri(const xml_node& node)
+        PUGI__FN const char_t* namespace_uri(xml_node node)
         {
                 namespace_uri_predicate pred = node.name();
@@ -6528 +8140 @@
         }
 
-        PUGI__FN const char_t* namespace_uri(const xml_attribute& attr, const xml_node& parent)
+        PUGI__FN const char_t* namespace_uri(xml_attribute attr, xml_node parent)
         {
                 namespace_uri_predicate pred = attr.name();
@@ -6554 +8166 @@
         }
 
-        PUGI__FN void normalize_space(char_t* buffer)
+        PUGI__FN char_t* normalize_space(char_t* buffer)
         {
                 char_t* write = buffer;
@@ -6578 +8190 @@
                 // zero-terminate
                 *write = 0;
-        }
-
-        PUGI__FN void translate(char_t* buffer, const char_t* from, const char_t* to)
-        {
-                size_t to_length = strlength(to);
-
+
+                return write;
+        }
+
+        PUGI__FN char_t* translate(char_t* buffer, const char_t* from, const char_t* to, size_t to_length)
+        {
                 char_t* write = buffer;
 
@@ -6600 +8212 @@
                 // zero-terminate
                 *write = 0;
+
+                return write;
+        }
+
+        PUGI__FN unsigned char* translate_table_generate(xpath_allocator* alloc, const char_t* from, const char_t* to)
+        {
+                unsigned char table[128] = {0};
+
+                while (*from)
+                {
+                        unsigned int fc = static_cast<unsigned int>(*from);
+                        unsigned int tc = static_cast<unsigned int>(*to);
+
+                        if (fc >= 128 || tc >= 128)
+                                return 0;
+
+                        // code=128 means "skip character"
+                        if (!table[fc])
+                                table[fc] = static_cast<unsigned char>(tc ? tc : 128);
+
+                        from++;
+                        if (tc) to++;
+                }
+
+                for (int i = 0; i < 128; ++i)
+                        if (!table[i])
+                                table[i] = static_cast<unsigned char>(i);
+
+                void* result = alloc->allocate_nothrow(sizeof(table));
+
+                if (result)
+                {
+                        memcpy(result, table, sizeof(table));
+                }
+
+                return static_cast<unsigned char*>(result);
+        }
+
+        PUGI__FN char_t* translate_table(char_t* buffer, const unsigned char* table)
+        {
+                char_t* write = buffer;
+
+                while (*buffer)
+                {
+                        char_t ch = *buffer++;
+                        unsigned int index = static_cast<unsigned int>(ch);
+
+                        if (index < 128)
+                        {
+                                unsigned char code = table[index];
+
+                                // code=128 means "skip character" (table size is 128 so 128 can be a special value)
+                                // this code skips these characters without extra branches
+                                *write = static_cast<char_t>(code);
+                                write += 1 - (code >> 7);
+                        }
+                        else
+                        {
+                                *write++ = ch;
+                        }
+                }
+
+                // zero-terminate
+                *write = 0;
+
+                return write;
+        }
+
+        inline bool is_xpath_attribute(const char_t* name)
+        {
+                return !(starts_with(name, PUGIXML_TEXT("xmlns")) && (name[5] == 0 || name[5] == ':'));
         }
 
         struct xpath_variable_boolean: xpath_variable
         {
-                xpath_variable_boolean(): value(false)
+                xpath_variable_boolean(): xpath_variable(xpath_type_boolean), value(false)
                 {
                 }
@@ -6614 +8297 @@
         struct xpath_variable_number: xpath_variable
         {
-                xpath_variable_number(): value(0)
+                xpath_variable_number(): xpath_variable(xpath_type_number), value(0)
                 {
                 }
@@ -6624 +8307 @@
         struct xpath_variable_string: xpath_variable
         {
-                xpath_variable_string(): value(0)
+                xpath_variable_string(): xpath_variable(xpath_type_string), value(0)
                 {
                 }
@@ -6639 +8322 @@
         struct xpath_variable_node_set: xpath_variable
         {
+                xpath_variable_node_set(): xpath_variable(xpath_type_node_set)
+                {
+                }
+
                 xpath_node_set value;
                 char_t name[1];
@@ -6732 +8419 @@
         }
 
-        PUGI__FN xpath_variable* get_variable(xpath_variable_set* set, const char_t* begin, const char_t* end)
-        {
-                char_t buffer[32];
-
+        PUGI__FN bool copy_xpath_variable(xpath_variable* lhs, const xpath_variable* rhs)
+        {
+                switch (rhs->type())
+                {
+                case xpath_type_node_set:
+                        return lhs->set(static_cast<const xpath_variable_node_set*>(rhs)->value);
+
+                case xpath_type_number:
+                        return lhs->set(static_cast<const xpath_variable_number*>(rhs)->value);
+
+                case xpath_type_string:
+                        return lhs->set(static_cast<const xpath_variable_string*>(rhs)->value);
+
+                case xpath_type_boolean:
+                        return lhs->set(static_cast<const xpath_variable_boolean*>(rhs)->value);
+
+                default:
+                        assert(!"Invalid variable type");
+                        return false;
+                }
+        }
+
+        PUGI__FN bool get_variable_scratch(char_t (&buffer)[32], xpath_variable_set* set, const char_t* begin, const char_t* end, xpath_variable** out_result)
+        {
                 size_t length = static_cast<size_t>(end - begin);
                 char_t* scratch = buffer;
@@ -6743 +8450 @@
                         // need to make dummy on-heap copy
                         scratch = static_cast<char_t*>(xml_memory::allocate((length + 1) * sizeof(char_t)));
-                        if (!scratch) return 0;
+                        if (!scratch) return false;
                 }
 
@@ -6750 +8457 @@
                 scratch[length] = 0;
 
-                xpath_variable* result = set->get(scratch);
+                *out_result = set->get(scratch);
 
                 // free dummy buffer
                 if (scratch != buffer) xml_memory::deallocate(scratch);
 
-                return result;
+                return true;
         }
 PUGI__NS_END
@@ -6761 +8468 @@
 // Internal node set class
 PUGI__NS_BEGIN
+        PUGI__FN xpath_node_set::type_t xpath_get_order(const xpath_node* begin, const xpath_node* end)
+        {
+                if (end - begin < 2)
+                        return xpath_node_set::type_sorted;
+
+                document_order_comparator cmp;
+
+                bool first = cmp(begin[0], begin[1]);
+
+                for (const xpath_node* it = begin + 1; it + 1 < end; ++it)
+                        if (cmp(it[0], it[1]) != first)
+                                return xpath_node_set::type_unsorted;
+
+                return first ? xpath_node_set::type_sorted : xpath_node_set::type_sorted_reverse;
+        }
+
         PUGI__FN xpath_node_set::type_t xpath_sort(xpath_node* begin, xpath_node* end, xpath_node_set::type_t type, bool rev)
         {
@@ -6767 +8490 @@
                 if (type == xpath_node_set::type_unsorted)
                 {
-                        sort(begin, end, document_order_comparator());
-
-                        type = xpath_node_set::type_sorted;
+                        xpath_node_set::type_t sorted = xpath_get_order(begin, end);
+
+                        if (sorted == xpath_node_set::type_unsorted)
+                        {
+                                sort(begin, end, document_order_comparator());
+
+                                type = xpath_node_set::type_sorted;
+                        }
+                        else
+                                type = sorted;
                 }
                 
@@ -6836 +8566 @@
                 }
 
+                void push_back_grow(const xpath_node& node, xpath_allocator* alloc);
+
                 void push_back(const xpath_node& node, xpath_allocator* alloc)
                 {
-                        if (_end == _eos)
-                        {
-                                size_t capacity = static_cast<size_t>(_eos - _begin);
-
-                                // get new capacity (1.5x rule)
-                                size_t new_capacity = capacity + capacity / 2 + 1;
-
-                                // reallocate the old array or allocate a new one
-                                xpath_node* data = static_cast<xpath_node*>(alloc->reallocate(_begin, capacity * sizeof(xpath_node), new_capacity * sizeof(xpath_node)));
-                                assert(data);
-
-                                // finalize
-                                _begin = data;
-                                _end = data + capacity;
-                                _eos = data + new_capacity;
-                        }
-
-                        *_end++ = node;
+                        if (_end != _eos)
+                                *_end++ = node;
+                        else
+                                push_back_grow(node, alloc);
                 }
 
                 void append(const xpath_node* begin_, const xpath_node* end_, xpath_allocator* alloc)
                 {
+                        if (begin_ == end_) return;
+
                         size_t size_ = static_cast<size_t>(_end - _begin);
                         size_t capacity = static_cast<size_t>(_eos - _begin);
@@ -6910 +8630 @@
                 }
         };
+
+        PUGI__FN_NO_INLINE void xpath_node_set_raw::push_back_grow(const xpath_node& node, xpath_allocator* alloc)
+        {
+                size_t capacity = static_cast<size_t>(_eos - _begin);
+
+                // get new capacity (1.5x rule)
+                size_t new_capacity = capacity + capacity / 2 + 1;
+
+                // reallocate the old array or allocate a new one
+                xpath_node* data = static_cast<xpath_node*>(alloc->reallocate(_begin, capacity * sizeof(xpath_node), new_capacity * sizeof(xpath_node)));
+                assert(data);
+
+                // finalize
+                _begin = data;
+                _end = data + capacity;
+                _eos = data + new_capacity;
+
+                // push
+                *_end++ = node;
+        }
 PUGI__NS_END
 
@@ -7279 +9019 @@
         enum ast_type_t
         {
+                ast_unknown,
                 ast_op_or,                                              // left or right
                 ast_op_and,                                             // left and right
@@ -7296 +9037 @@
                 ast_predicate,                                  // apply predicate to set; next points to next predicate
                 ast_filter,                                             // select * from left where right
-                ast_filter_posinv,                              // select * from left where right; proximity position invariant
                 ast_string_constant,                    // string constant
                 ast_number_constant,                    // number constant
@@ -7336 +9076 @@
                 ast_func_round,                                 // round(left)
                 ast_step,                                               // process set left with step
-                ast_step_root                                   // select root node
+                ast_step_root,                                  // select root node
+
+                ast_opt_translate_table,                // translate(left, right, third) where right/third are constants
+                ast_opt_compare_attribute               // @name = 'string'
         };
 
@@ -7369 +9112 @@
         };
 
+        enum predicate_t
+        {
+                predicate_default,
+                predicate_posinv,
+                predicate_constant,
+                predicate_constant_one
+        };
+
+        enum nodeset_eval_t
+        {
+                nodeset_eval_all,
+                nodeset_eval_any,
+                nodeset_eval_first
+        };
+
         template <axis_t N> struct axis_to_type
         {
@@ -7383 +9141 @@
                 char _rettype;
 
-                // for ast_step / ast_predicate
+                // for ast_step
                 char _axis;
+
+                // for ast_step/ast_predicate/ast_filter
                 char _test;
 
@@ -7402 +9162 @@
                         // node test for ast_step (node name/namespace/node type/pi target)
                         const char_t* nodetest;
+                        // table for ast_opt_translate_table
+                        const unsigned char* table;
                 } _data;
 
@@ -7431 +9193 @@
                                 xpath_allocator_capture cr(stack.result);
 
-                                xpath_node_set_raw ls = lhs->eval_node_set(c, stack);
-                                xpath_node_set_raw rs = rhs->eval_node_set(c, stack);
+                                xpath_node_set_raw ls = lhs->eval_node_set(c, stack, nodeset_eval_all);
+                                xpath_node_set_raw rs = rhs->eval_node_set(c, stack, nodeset_eval_all);
 
                                 for (const xpath_node* li = ls.begin(); li != ls.end(); ++li)
@@ -7460 +9222 @@
 
                                         double l = lhs->eval_number(c, stack);
-                                        xpath_node_set_raw rs = rhs->eval_node_set(c, stack);
+                                        xpath_node_set_raw rs = rhs->eval_node_set(c, stack, nodeset_eval_all);
 
                                         for (const xpath_node* ri = rs.begin(); ri != rs.end(); ++ri)
@@ -7477 +9239 @@
 
                                         xpath_string l = lhs->eval_string(c, stack);
-                                        xpath_node_set_raw rs = rhs->eval_node_set(c, stack);
+                                        xpath_node_set_raw rs = rhs->eval_node_set(c, stack, nodeset_eval_all);
 
                                         for (const xpath_node* ri = rs.begin(); ri != rs.end(); ++ri)
@@ -7495 +9257 @@
                 }
 
+                static bool eval_once(xpath_node_set::type_t type, nodeset_eval_t eval)
+                {
+                        return type == xpath_node_set::type_sorted ? eval != nodeset_eval_all : eval == nodeset_eval_any;
+                }
+
                 template <class Comp> static bool compare_rel(xpath_ast_node* lhs, xpath_ast_node* rhs, const xpath_context& c, const xpath_stack& stack, const Comp& comp)
                 {
@@ -7505 +9272 @@
                                 xpath_allocator_capture cr(stack.result);
 
-                                xpath_node_set_raw ls = lhs->eval_node_set(c, stack);
-                                xpath_node_set_raw rs = rhs->eval_node_set(c, stack);
+                                xpath_node_set_raw ls = lhs->eval_node_set(c, stack, nodeset_eval_all);
+                                xpath_node_set_raw rs = rhs->eval_node_set(c, stack, nodeset_eval_all);
 
                                 for (const xpath_node* li = ls.begin(); li != ls.end(); ++li)
@@ -7530 +9297 @@
 
                                 double l = lhs->eval_number(c, stack);
-                                xpath_node_set_raw rs = rhs->eval_node_set(c, stack);
+                                xpath_node_set_raw rs = rhs->eval_node_set(c, stack, nodeset_eval_all);
 
                                 for (const xpath_node* ri = rs.begin(); ri != rs.end(); ++ri)
@@ -7546 +9313 @@
                                 xpath_allocator_capture cr(stack.result);
 
-                                xpath_node_set_raw ls = lhs->eval_node_set(c, stack);
+                                xpath_node_set_raw ls = lhs->eval_node_set(c, stack, nodeset_eval_all);
                                 double r = rhs->eval_number(c, stack);
 
@@ -7566 +9333 @@
                 }
 
-                void apply_predicate(xpath_node_set_raw& ns, size_t first, xpath_ast_node* expr, const xpath_stack& stack)
+                static void apply_predicate_boolean(xpath_node_set_raw& ns, size_t first, xpath_ast_node* expr, const xpath_stack& stack, bool once)
                 {
                         assert(ns.size() >= first);
+                        assert(expr->rettype() != xpath_type_number);
 
                         size_t i = 1;
                         size_t size = ns.size() - first;
-                                
+
                         xpath_node* last = ns.begin() + first;
-                                
+
                         // remove_if... or well, sort of
                         for (xpath_node* it = last; it != ns.end(); ++it, ++i)
                         {
                                 xpath_context c(*it, i, size);
-                        
-                                if (expr->rettype() == xpath_type_number)
-                                {
-                                        if (expr->eval_number(c, stack) == i)
-                                                *last++ = *it;
-                                }
-                                else if (expr->eval_boolean(c, stack))
+
+                                if (expr->eval_boolean(c, stack))
+                                {
                                         *last++ = *it;
-                        }
-                        
+
+                                        if (once) break;
+                                }
+                        }
+
                         ns.truncate(last);
                 }
 
-                void apply_predicates(xpath_node_set_raw& ns, size_t first, const xpath_stack& stack)
+                static void apply_predicate_number(xpath_node_set_raw& ns, size_t first, xpath_ast_node* expr, const xpath_stack& stack, bool once)
+                {
+                        assert(ns.size() >= first);
+                        assert(expr->rettype() == xpath_type_number);
+
+                        size_t i = 1;
+                        size_t size = ns.size() - first;
+
+                        xpath_node* last = ns.begin() + first;
+
+                        // remove_if... or well, sort of
+                        for (xpath_node* it = last; it != ns.end(); ++it, ++i)
+                        {
+                                xpath_context c(*it, i, size);
+
+                                if (expr->eval_number(c, stack) == i)
+                                {
+                                        *last++ = *it;
+
+                                        if (once) break;
+                                }
+                        }
+
+                        ns.truncate(last);
+                }
+
+                static void apply_predicate_number_const(xpath_node_set_raw& ns, size_t first, xpath_ast_node* expr, const xpath_stack& stack)
+                {
+                        assert(ns.size() >= first);
+                        assert(expr->rettype() == xpath_type_number);
+
+                        size_t size = ns.size() - first;
+
+                        xpath_node* last = ns.begin() + first;
+
+                        xpath_context c(xpath_node(), 1, size);
+
+                        double er = expr->eval_number(c, stack);
+
+                        if (er >= 1.0 && er <= size)
+                        {
+                                size_t eri = static_cast<size_t>(er);
+
+                                if (er == eri)
+                                {
+                                        xpath_node r = last[eri - 1];
+
+                                        *last++ = r;
+                                }
+                        }
+
+                        ns.truncate(last);
+                }
+
+                void apply_predicate(xpath_node_set_raw& ns, size_t first, const xpath_stack& stack, bool once)
                 {
                         if (ns.size() == first) return;
-                        
+
+                        assert(_type == ast_filter || _type == ast_predicate);
+
+                        if (_test == predicate_constant || _test == predicate_constant_one)
+                                apply_predicate_number_const(ns, first, _right, stack);
+                        else if (_right->rettype() == xpath_type_number)
+                                apply_predicate_number(ns, first, _right, stack, once);
+                        else
+                                apply_predicate_boolean(ns, first, _right, stack, once);
+                }
+
+                void apply_predicates(xpath_node_set_raw& ns, size_t first, const xpath_stack& stack, nodeset_eval_t eval)
+                {
+                        if (ns.size() == first) return;
+
+                        bool last_once = eval_once(ns.type(), eval);
+
                         for (xpath_ast_node* pred = _right; pred; pred = pred->_next)
-                        {
-                                apply_predicate(ns, first, pred->_left, stack);
-                        }
-                }
-
-                void step_push(xpath_node_set_raw& ns, const xml_attribute& a, const xml_node& parent, xpath_allocator* alloc)
-                {
-                        if (!a) return;
-
-                        const char_t* name = a.name();
-
-                        // There are no attribute nodes corresponding to attributes that declare namespaces
-                        // That is, "xmlns:..." or "xmlns"
-                        if (starts_with(name, PUGIXML_TEXT("xmlns")) && (name[5] == 0 || name[5] == ':')) return;
-                        
+                                pred->apply_predicate(ns, first, stack, !pred->_next && last_once);
+                }
+
+                bool step_push(xpath_node_set_raw& ns, xml_attribute_struct* a, xml_node_struct* parent, xpath_allocator* alloc)
+                {
+                        assert(a);
+
+                        const char_t* name = a->name ? a->name + 0 : PUGIXML_TEXT("");
+
                         switch (_test)
                         {
                         case nodetest_name:
-                                if (strequal(name, _data.nodetest)) ns.push_back(xpath_node(a, parent), alloc);
+                                if (strequal(name, _data.nodetest) && is_xpath_attribute(name))
+                                {
+                                        ns.push_back(xpath_node(xml_attribute(a), xml_node(parent)), alloc);
+                                        return true;
+                                }
                                 break;
                                 
                         case nodetest_type_node:
                         case nodetest_all:
-                                ns.push_back(xpath_node(a, parent), alloc);
+                                if (is_xpath_attribute(name))
+                                {
+                                        ns.push_back(xpath_node(xml_attribute(a), xml_node(parent)), alloc);
+                                        return true;
+                                }
                                 break;
                                 
                         case nodetest_all_in_namespace:
-                                if (starts_with(name, _data.nodetest))
-                                        ns.push_back(xpath_node(a, parent), alloc);
+                                if (starts_with(name, _data.nodetest) && is_xpath_attribute(name))
+                                {
+                                        ns.push_back(xpath_node(xml_attribute(a), xml_node(parent)), alloc);
+                                        return true;
+                                }
                                 break;
                         
@@ -7631 +9473 @@
                                 ;
                         }
+
+                        return false;
                 }
                 
-                void step_push(xpath_node_set_raw& ns, const xml_node& n, xpath_allocator* alloc)
-                {
-                        if (!n) return;
+                bool step_push(xpath_node_set_raw& ns, xml_node_struct* n, xpath_allocator* alloc)
+                {
+                        assert(n);
+
+                        xml_node_type type = PUGI__NODETYPE(n);
 
                         switch (_test)
                         {
                         case nodetest_name:
-                                if (n.type() == node_element && strequal(n.name(), _data.nodetest)) ns.push_back(n, alloc);
+                                if (type == node_element && n->name && strequal(n->name, _data.nodetest))
+                                {
+                                        ns.push_back(xml_node(n), alloc);
+                                        return true;
+                                }
                                 break;
                                 
                         case nodetest_type_node:
-                                ns.push_back(n, alloc);
-                                break;
+                                ns.push_back(xml_node(n), alloc);
+                                return true;
                                 
                         case nodetest_type_comment:
-                                if (n.type() == node_comment)
-                                        ns.push_back(n, alloc);
+                                if (type == node_comment)
+                                {
+                                        ns.push_back(xml_node(n), alloc);
+                                        return true;
+                                }
                                 break;
                                 
                         case nodetest_type_text:
-                                if (n.type() == node_pcdata || n.type() == node_cdata)
-                                        ns.push_back(n, alloc);
+                                if (type == node_pcdata || type == node_cdata)
+                                {
+                                        ns.push_back(xml_node(n), alloc);
+                                        return true;
+                                }
                                 break;
                                 
                         case nodetest_type_pi:
-                                if (n.type() == node_pi)
-                                        ns.push_back(n, alloc);
+                                if (type == node_pi)
+                                {
+                                        ns.push_back(xml_node(n), alloc);
+                                        return true;
+                                }
                                 break;
                                                                         
                         case nodetest_pi:
-                                if (n.type() == node_pi && strequal(n.name(), _data.nodetest))
-                                        ns.push_back(n, alloc);
+                                if (type == node_pi && n->name && strequal(n->name, _data.nodetest))
+                                {
+                                        ns.push_back(xml_node(n), alloc);
+                                        return true;
+                                }
                                 break;
                                 
                         case nodetest_all:
-                                if (n.type() == node_element)
-                                        ns.push_back(n, alloc);
+                                if (type == node_element)
+                                {
+                                        ns.push_back(xml_node(n), alloc);
+                                        return true;
+                                }
                                 break;
                                 
                         case nodetest_all_in_namespace:
-                                if (n.type() == node_element && starts_with(n.name(), _data.nodetest))
-                                        ns.push_back(n, alloc);
+                                if (type == node_element && n->name && starts_with(n->name, _data.nodetest))
+                                {
+                                        ns.push_back(xml_node(n), alloc);
+                                        return true;
+                                }
                                 break;
 
                         default:
                                 assert(!"Unknown axis");
-                        } 
-                }
-
-                template <class T> void step_fill(xpath_node_set_raw& ns, const xml_node& n, xpath_allocator* alloc, T)
+                        }
+
+                        return false;
+                }
+
+                template <class T> void step_fill(xpath_node_set_raw& ns, xml_node_struct* n, xpath_allocator* alloc, bool once, T)
                 {
                         const axis_t axis = T::axis;
@@ -7690 +9560 @@
                         case axis_attribute:
                         {
-                                for (xml_attribute a = n.first_attribute(); a; a = a.next_attribute())
-                                        step_push(ns, a, n, alloc);
+                                for (xml_attribute_struct* a = n->first_attribute; a; a = a->next_attribute)
+                                        if (step_push(ns, a, n, alloc) & once)
+                                                return;
                                 
                                 break;
@@ -7698 +9569 @@
                         case axis_child:
                         {
-                                for (xml_node c = n.first_child(); c; c = c.next_sibling())
-                                        step_push(ns, c, alloc);
+                                for (xml_node_struct* c = n->first_child; c; c = c->next_sibling)
+                                        if (step_push(ns, c, alloc) & once)
+                                                return;
                                         
                                 break;
@@ -7708 +9580 @@
                         {
                                 if (axis == axis_descendant_or_self)
-                                        step_push(ns, n, alloc);
+                                        if (step_push(ns, n, alloc) & once)
+                                                return;
                                         
-                                xml_node cur = n.first_child();
+                                xml_node_struct* cur = n->first_child;
                                 
-                                while (cur && cur != n)
-                                {
-                                        step_push(ns, cur, alloc);
+                                while (cur)
+                                {
+                                        if (step_push(ns, cur, alloc) & once)
+                                                return;
                                         
-                                        if (cur.first_child())
-                                                cur = cur.first_child();
-                                        else if (cur.next_sibling())
-                                                cur = cur.next_sibling();
+                                        if (cur->first_child)
+                                                cur = cur->first_child;
                                         else
                                         {
-                                                while (!cur.next_sibling() && cur != n)
-                                                        cur = cur.parent();
+                                                while (!cur->next_sibling)
+                                                {
+                                                        cur = cur->parent;
+
+                                                        if (cur == n) return;
+                                                }
                                         
-                                                if (cur != n) cur = cur.next_sibling();
+                                                cur = cur->next_sibling;
                                         }
                                 }
@@ -7734 +9610 @@
                         case axis_following_sibling:
                         {
-                                for (xml_node c = n.next_sibling(); c; c = c.next_sibling())
-                                        step_push(ns, c, alloc);
+                                for (xml_node_struct* c = n->next_sibling; c; c = c->next_sibling)
+                                        if (step_push(ns, c, alloc) & once)
+                                                return;
                                 
                                 break;
@@ -7742 +9619 @@
                         case axis_preceding_sibling:
                         {
-                                for (xml_node c = n.previous_sibling(); c; c = c.previous_sibling())
-                                        step_push(ns, c, alloc);
+                                for (xml_node_struct* c = n->prev_sibling_c; c->next_sibling; c = c->prev_sibling_c)
+                                        if (step_push(ns, c, alloc) & once)
+                                                return;
                                 
                                 break;
@@ -7750 +9628 @@
                         case axis_following:
                         {
-                                xml_node cur = n;
+                                xml_node_struct* cur = n;
 
                                 // exit from this node so that we don't include descendants
-                                while (cur && !cur.next_sibling()) cur = cur.parent();
-                                cur = cur.next_sibling();
-
-                                for (;;)
-                                {
-                                        step_push(ns, cur, alloc);
-
-                                        if (cur.first_child())
-                                                cur = cur.first_child();
-                                        else if (cur.next_sibling())
-                                                cur = cur.next_sibling();
+                                while (!cur->next_sibling)
+                                {
+                                        cur = cur->parent;
+
+                                        if (!cur) return;
+                                }
+
+                                cur = cur->next_sibling;
+
+                                while (cur)
+                                {
+                                        if (step_push(ns, cur, alloc) & once)
+                                                return;
+
+                                        if (cur->first_child)
+                                                cur = cur->first_child;
                                         else
                                         {
-                                                while (cur && !cur.next_sibling()) cur = cur.parent();
-                                                cur = cur.next_sibling();
-
-                                                if (!cur) break;
+                                                while (!cur->next_sibling)
+                                                {
+                                                        cur = cur->parent;
+
+                                                        if (!cur) return;
+                                                }
+
+                                                cur = cur->next_sibling;
                                         }
                                 }
@@ -7778 +9665 @@
                         case axis_preceding:
                         {
-                                xml_node cur = n;
-
-                                while (cur && !cur.previous_sibling()) cur = cur.parent();
-                                cur = cur.previous_sibling();
-
-                                for (;;)
-                                {
-                                        if (cur.last_child())
-                                                cur = cur.last_child();
+                                xml_node_struct* cur = n;
+
+                                // exit from this node so that we don't include descendants
+                                while (!cur->prev_sibling_c->next_sibling)
+                                {
+                                        cur = cur->parent;
+
+                                        if (!cur) return;
+                                }
+
+                                cur = cur->prev_sibling_c;
+
+                                while (cur)
+                                {
+                                        if (cur->first_child)
+                                                cur = cur->first_child->prev_sibling_c;
                                         else
                                         {
                                                 // leaf node, can't be ancestor
-                                                step_push(ns, cur, alloc);
-
-                                                if (cur.previous_sibling())
-                                                        cur = cur.previous_sibling();
-                                                else
+                                                if (step_push(ns, cur, alloc) & once)
+                                                        return;
+
+                                                while (!cur->prev_sibling_c->next_sibling)
                                                 {
-                                                        do 
-                                                        {
-                                                                cur = cur.parent();
-                                                                if (!cur) break;
-
-                                                                if (!node_is_ancestor(cur, n)) step_push(ns, cur, alloc);
-                                                        }
-                                                        while (!cur.previous_sibling());
-
-                                                        cur = cur.previous_sibling();
-
-                                                        if (!cur) break;
+                                                        cur = cur->parent;
+
+                                                        if (!cur) return;
+
+                                                        if (!node_is_ancestor(cur, n))
+                                                                if (step_push(ns, cur, alloc) & once)
+                                                                        return;
                                                 }
+
+                                                cur = cur->prev_sibling_c;
                                         }
                                 }
@@ -7819 +9709 @@
                         {
                                 if (axis == axis_ancestor_or_self)
-                                        step_push(ns, n, alloc);
-
-                                xml_node cur = n.parent();
+                                        if (step_push(ns, n, alloc) & once)
+                                                return;
+
+                                xml_node_struct* cur = n->parent;
                                 
                                 while (cur)
                                 {
-                                        step_push(ns, cur, alloc);
+                                        if (step_push(ns, cur, alloc) & once)
+                                                return;
                                         
-                                        cur = cur.parent();
+                                        cur = cur->parent;
                                 }
                                 
@@ -7842 +9734 @@
                         case axis_parent:
                         {
-                                if (n.parent()) step_push(ns, n.parent(), alloc);
+                                if (n->parent)
+                                        step_push(ns, n->parent, alloc);
 
                                 break;
@@ -7852 +9745 @@
                 }
                 
-                template <class T> void step_fill(xpath_node_set_raw& ns, const xml_attribute& a, const xml_node& p, xpath_allocator* alloc, T v)
+                template <class T> void step_fill(xpath_node_set_raw& ns, xml_attribute_struct* a, xml_node_struct* p, xpath_allocator* alloc, bool once, T v)
                 {
                         const axis_t axis = T::axis;
@@ -7862 +9755 @@
                         {
                                 if (axis == axis_ancestor_or_self && _test == nodetest_type_node) // reject attributes based on principal node type test
-                                        step_push(ns, a, p, alloc);
-
-                                xml_node cur = p;
+                                        if (step_push(ns, a, p, alloc) & once)
+                                                return;
+
+                                xml_node_struct* cur = p;
                                 
                                 while (cur)
                                 {
-                                        step_push(ns, cur, alloc);
+                                        if (step_push(ns, cur, alloc) & once)
+                                                return;
                                         
-                                        cur = cur.parent();
+                                        cur = cur->parent;
                                 }
                                 
@@ -7887 +9782 @@
                         case axis_following:
                         {
-                                xml_node cur = p;
+                                xml_node_struct* cur = p;
                                 
-                                for (;;)
-                                {
-                                        if (cur.first_child())
-                                                cur = cur.first_child();
-                                        else if (cur.next_sibling())
-                                                cur = cur.next_sibling();
+                                while (cur)
+                                {
+                                        if (cur->first_child)
+                                                cur = cur->first_child;
                                         else
                                         {
-                                                while (cur && !cur.next_sibling()) cur = cur.parent();
-                                                cur = cur.next_sibling();
-                                                
-                                                if (!cur) break;
+                                                while (!cur->next_sibling)
+                                                {
+                                                        cur = cur->parent;
+
+                                                        if (!cur) return;
+                                                }
+
+                                                cur = cur->next_sibling;
                                         }
 
-                                        step_push(ns, cur, alloc);
+                                        if (step_push(ns, cur, alloc) & once)
+                                                return;
                                 }
 
@@ -7919 +9817 @@
                         {
                                 // preceding:: axis does not include attribute nodes and attribute ancestors (they are the same as parent's ancestors), so we can reuse node preceding
-                                step_fill(ns, p, alloc, v);
+                                step_fill(ns, p, alloc, once, v);
                                 break;
                         }
@@ -7927 +9825 @@
                         }
                 }
-                
-                template <class T> xpath_node_set_raw step_do(const xpath_context& c, const xpath_stack& stack, T v)
+
+                template <class T> void step_fill(xpath_node_set_raw& ns, const xpath_node& xn, xpath_allocator* alloc, bool once, T v)
                 {
                         const axis_t axis = T::axis;
-                        bool attributes = (axis == axis_ancestor || axis == axis_ancestor_or_self || axis == axis_descendant_or_self || axis == axis_following || axis == axis_parent || axis == axis_preceding || axis == axis_self);
+                        const bool axis_has_attributes = (axis == axis_ancestor || axis == axis_ancestor_or_self || axis == axis_descendant_or_self || axis == axis_following || axis == axis_parent || axis == axis_preceding || axis == axis_self);
+
+                        if (xn.node())
+                                step_fill(ns, xn.node().internal_object(), alloc, once, v);
+                        else if (axis_has_attributes && xn.attribute() && xn.parent())
+                                step_fill(ns, xn.attribute().internal_object(), xn.parent().internal_object(), alloc, once, v);
+                }
+
+                template <class T> xpath_node_set_raw step_do(const xpath_context& c, const xpath_stack& stack, nodeset_eval_t eval, T v)
+                {
+                        const axis_t axis = T::axis;
+                        const bool axis_reverse = (axis == axis_ancestor || axis == axis_ancestor_or_self || axis == axis_preceding || axis == axis_preceding_sibling);
+                        const xpath_node_set::type_t axis_type = axis_reverse ? xpath_node_set::type_sorted_reverse : xpath_node_set::type_sorted;
+
+                        bool once =
+                                (axis == axis_attribute && _test == nodetest_name) ||
+                                (!_right && eval_once(axis_type, eval)) ||
+                                (_right && !_right->_next && _right->_test == predicate_constant_one);
 
                         xpath_node_set_raw ns;
-                        ns.set_type((axis == axis_ancestor || axis == axis_ancestor_or_self || axis == axis_preceding || axis == axis_preceding_sibling) ? xpath_node_set::type_sorted_reverse : xpath_node_set::type_sorted);
+                        ns.set_type(axis_type);
 
                         if (_left)
                         {
-                                xpath_node_set_raw s = _left->eval_node_set(c, stack);
+                                xpath_node_set_raw s = _left->eval_node_set(c, stack, nodeset_eval_all);
 
                                 // self axis preserves the original order
@@ -7950 +9865 @@
                                         if (axis != axis_self && size != 0) ns.set_type(xpath_node_set::type_unsorted);
                                         
-                                        if (it->node())
-                                                step_fill(ns, it->node(), stack.result, v);
-                                        else if (attributes)
-                                                step_fill(ns, it->attribute(), it->parent(), stack.result, v);
-                                                
-                                        apply_predicates(ns, size, stack);
+                                        step_fill(ns, *it, stack.result, once, v);
+                                        if (_right) apply_predicates(ns, size, stack, eval);
                                 }
                         }
                         else
                         {
-                                if (c.n.node())
-                                        step_fill(ns, c.n.node(), stack.result, v);
-                                else if (attributes)
-                                        step_fill(ns, c.n.attribute(), c.n.parent(), stack.result, v);
-                                
-                                apply_predicates(ns, 0, stack);
+                                step_fill(ns, c.n, stack.result, once, v);
+                                if (_right) apply_predicates(ns, 0, stack, eval);
                         }
 
@@ -8006 +9913 @@
                         _type(static_cast<char>(type)), _rettype(xpath_type_node_set), _axis(static_cast<char>(axis)), _test(static_cast<char>(test)), _left(left), _right(0), _next(0)
                 {
+                        assert(type == ast_step);
                         _data.nodetest = contents;
+                }
+
+                xpath_ast_node(ast_type_t type, xpath_ast_node* left, xpath_ast_node* right, predicate_t test):
+                        _type(static_cast<char>(type)), _rettype(xpath_type_node_set), _axis(0), _test(static_cast<char>(test)), _left(left), _right(right), _next(0)
+                {
+                        assert(type == ast_filter || type == ast_predicate);
                 }
 
@@ -8109 +10023 @@
                         }
 
+                        case ast_opt_compare_attribute:
+                        {
+                                const char_t* value = (_right->_type == ast_string_constant) ? _right->_data.string : _right->_data.variable->get_string();
+
+                                xml_attribute attr = c.n.node().attribute(_left->_data.nodetest);
+
+                                return attr && strequal(attr.value(), value) && is_xpath_attribute(attr.name());
+                        }
+
                         case ast_variable:
                         {
@@ -8137 +10060 @@
                                         xpath_allocator_capture cr(stack.result);
 
-                                        return !eval_node_set(c, stack).empty();
+                                        return !eval_node_set(c, stack, nodeset_eval_any).empty();
                                 }
 
@@ -8183 +10106 @@
                                 xpath_allocator_capture cr(stack.result);
 
-                                return static_cast<double>(_left->eval_node_set(c, stack).size());
+                                return static_cast<double>(_left->eval_node_set(c, stack, nodeset_eval_all).size());
                         }
                         
@@ -8216 +10139 @@
                                 double r = 0;
                                 
-                                xpath_node_set_raw ns = _left->eval_node_set(c, stack);
+                                xpath_node_set_raw ns = _left->eval_node_set(c, stack, nodeset_eval_all);
                                 
                                 for (const xpath_node* it = ns.begin(); it != ns.end(); ++it)
@@ -8331 +10254 @@
                         *ri = 0;
 
-                        return xpath_string(result, true);
+                        return xpath_string::from_heap_preallocated(result, ri);
                 }
 
@@ -8339 +10262 @@
                         {
                         case ast_string_constant:
-                                return xpath_string_const(_data.string);
+                                return xpath_string::from_const(_data.string);
                         
                         case ast_func_local_name_0:
@@ -8345 +10268 @@
                                 xpath_node na = c.n;
                                 
-                                return xpath_string_const(local_name(na));
+                                return xpath_string::from_const(local_name(na));
                         }
 
@@ -8352 +10275 @@
                                 xpath_allocator_capture cr(stack.result);
 
-                                xpath_node_set_raw ns = _left->eval_node_set(c, stack);
+                                xpath_node_set_raw ns = _left->eval_node_set(c, stack, nodeset_eval_first);
                                 xpath_node na = ns.first();
                                 
-                                return xpath_string_const(local_name(na));
+                                return xpath_string::from_const(local_name(na));
                         }
 
@@ -8362 +10285 @@
                                 xpath_node na = c.n;
                                 
-                                return xpath_string_const(qualified_name(na));
+                                return xpath_string::from_const(qualified_name(na));
                         }
 
@@ -8369 +10292 @@
                                 xpath_allocator_capture cr(stack.result);
 
-                                xpath_node_set_raw ns = _left->eval_node_set(c, stack);
+                                xpath_node_set_raw ns = _left->eval_node_set(c, stack, nodeset_eval_first);
                                 xpath_node na = ns.first();
                                 
-                                return xpath_string_const(qualified_name(na));
+                                return xpath_string::from_const(qualified_name(na));
                         }
 
@@ -8379 +10302 @@
                                 xpath_node na = c.n;
                                 
-                                return xpath_string_const(namespace_uri(na));
+                                return xpath_string::from_const(namespace_uri(na));
                         }
 
@@ -8386 +10309 @@
                                 xpath_allocator_capture cr(stack.result);
 
-                                xpath_node_set_raw ns = _left->eval_node_set(c, stack);
+                                xpath_node_set_raw ns = _left->eval_node_set(c, stack, nodeset_eval_first);
                                 xpath_node na = ns.first();
                                 
-                                return xpath_string_const(namespace_uri(na));
+                                return xpath_string::from_const(namespace_uri(na));
                         }
 
@@ -8412 +10335 @@
                                 const char_t* pos = find_substring(s.c_str(), p.c_str());
                                 
-                                return pos ? xpath_string(s.c_str(), pos, stack.result) : xpath_string();
+                                return pos ? xpath_string::from_heap(s.c_str(), pos, stack.result) : xpath_string();
                         }
                         
@@ -8427 +10350 @@
                                 if (!pos) return xpath_string();
 
-                                const char_t* result = pos + p.length();
-
-                                return s.uses_heap() ? xpath_string(result, stack.result) : xpath_string_const(result);
+                                const char_t* rbegin = pos + p.length();
+                                const char_t* rend = s.c_str() + s.length();
+
+                                return s.uses_heap() ? xpath_string::from_heap(rbegin, rend, stack.result) : xpath_string::from_const(rbegin);
                         }
 
@@ -8450 +10374 @@
 
                                 const char_t* rbegin = s.c_str() + (pos - 1);
+                                const char_t* rend = s.c_str() + s.length();
                                 
-                                return s.uses_heap() ? xpath_string(rbegin, stack.result) : xpath_string_const(rbegin);
+                                return s.uses_heap() ? xpath_string::from_heap(rbegin, rend, stack.result) : xpath_string::from_const(rbegin);
                         }
                         
@@ -8478 +10403 @@
                                 const char_t* rend = s.c_str() + (end - 1);
 
-                                return (end == s_length + 1 && !s.uses_heap()) ? xpath_string_const(rbegin) : xpath_string(rbegin, rend, stack.result);
+                                return (end == s_length + 1 && !s.uses_heap()) ? xpath_string::from_const(rbegin) : xpath_string::from_heap(rbegin, rend, stack.result);
                         }
 
@@ -8485 +10410 @@
                                 xpath_string s = string_value(c.n, stack.result);
 
-                                normalize_space(s.data(stack.result));
-
-                                return s;
+                                char_t* begin = s.data(stack.result);
+                                char_t* end = normalize_space(begin);
+
+                                return xpath_string::from_heap_preallocated(begin, end);
                         }
 
@@ -8494 +10420 @@
                                 xpath_string s = _left->eval_string(c, stack);
 
-                                normalize_space(s.data(stack.result));
+                                char_t* begin = s.data(stack.result);
+                                char_t* end = normalize_space(begin);
                         
-                                return s;
+                                return xpath_string::from_heap_preallocated(begin, end);
                         }
 
@@ -8509 +10436 @@
                                 xpath_string to = _right->_next->eval_string(c, swapped_stack);
 
-                                translate(s.data(stack.result), from.c_str(), to.c_str());
-
-                                return s;
+                                char_t* begin = s.data(stack.result);
+                                char_t* end = translate(begin, from.c_str(), to.c_str(), to.length());
+
+                                return xpath_string::from_heap_preallocated(begin, end);
+                        }
+
+                        case ast_opt_translate_table:
+                        {
+                                xpath_string s = _left->eval_string(c, stack);
+
+                                char_t* begin = s.data(stack.result);
+                                char_t* end = translate_table(begin, _data.table);
+
+                                return xpath_string::from_heap_preallocated(begin, end);
                         }
 
@@ -8519 +10457 @@
 
                                 if (_rettype == xpath_type_string)
-                                        return xpath_string_const(_data.variable->get_string());
+                                        return xpath_string::from_const(_data.variable->get_string());
 
                                 // fallthrough to type conversion
@@ -8529 +10467 @@
                                 {
                                 case xpath_type_boolean:
-                                        return xpath_string_const(eval_boolean(c, stack) ? PUGIXML_TEXT("true") : PUGIXML_TEXT("false"));
+                                        return xpath_string::from_const(eval_boolean(c, stack) ? PUGIXML_TEXT("true") : PUGIXML_TEXT("false"));
                                         
                                 case xpath_type_number:
@@ -8540 +10478 @@
                                         xpath_stack swapped_stack = {stack.temp, stack.result};
 
-                                        xpath_node_set_raw ns = eval_node_set(c, swapped_stack);
+                                        xpath_node_set_raw ns = eval_node_set(c, swapped_stack, nodeset_eval_first);
                                         return ns.empty() ? xpath_string() : string_value(ns.first(), stack.result);
                                 }
@@ -8552 +10490 @@
                 }
 
-                xpath_node_set_raw eval_node_set(const xpath_context& c, const xpath_stack& stack)
+                xpath_node_set_raw eval_node_set(const xpath_context& c, const xpath_stack& stack, nodeset_eval_t eval)
                 {
                         switch (_type)
@@ -8562 +10500 @@
                                 xpath_stack swapped_stack = {stack.temp, stack.result};
 
-                                xpath_node_set_raw ls = _left->eval_node_set(c, swapped_stack);
-                                xpath_node_set_raw rs = _right->eval_node_set(c, stack);
-                                
+                                xpath_node_set_raw ls = _left->eval_node_set(c, swapped_stack, eval);
+                                xpath_node_set_raw rs = _right->eval_node_set(c, stack, eval);
+
                                 // we can optimize merging two sorted sets, but this is a very rare operation, so don't bother
                                 rs.set_type(xpath_node_set::type_unsorted);
@@ -8570 +10508 @@
                                 rs.append(ls.begin(), ls.end(), stack.result);
                                 rs.remove_duplicates();
-                                
+
                                 return rs;
                         }
 
                         case ast_filter:
-                        case ast_filter_posinv:
-                        {
-                                xpath_node_set_raw set = _left->eval_node_set(c, stack);
+                        {
+                                xpath_node_set_raw set = _left->eval_node_set(c, stack, _test == predicate_constant_one ? nodeset_eval_first : nodeset_eval_all);
 
                                 // either expression is a number or it contains position() call; sort by document order
-                                if (_type == ast_filter) set.sort_do();
-
-                                apply_predicate(set, 0, _right, stack);
+                                if (_test != predicate_posinv) set.sort_do();
+
+                                bool once = eval_once(set.type(), eval);
+
+                                apply_predicate(set, 0, stack, once);
                         
                                 return set;
@@ -8595 +10534 @@
                                 {
                                 case axis_ancestor:
-                                        return step_do(c, stack, axis_to_type<axis_ancestor>());
+                                        return step_do(c, stack, eval, axis_to_type<axis_ancestor>());
                                         
                                 case axis_ancestor_or_self:
-                                        return step_do(c, stack, axis_to_type<axis_ancestor_or_self>());
+                                        return step_do(c, stack, eval, axis_to_type<axis_ancestor_or_self>());
 
                                 case axis_attribute:
-                                        return step_do(c, stack, axis_to_type<axis_attribute>());
+                                        return step_do(c, stack, eval, axis_to_type<axis_attribute>());
 
                                 case axis_child:
-                                        return step_do(c, stack, axis_to_type<axis_child>());
+                                        return step_do(c, stack, eval, axis_to_type<axis_child>());
                                 
                                 case axis_descendant:
-                                        return step_do(c, stack, axis_to_type<axis_descendant>());
+                                        return step_do(c, stack, eval, axis_to_type<axis_descendant>());
 
                                 case axis_descendant_or_self:
-                                        return step_do(c, stack, axis_to_type<axis_descendant_or_self>());
+                                        return step_do(c, stack, eval, axis_to_type<axis_descendant_or_self>());
 
                                 case axis_following:
-                                        return step_do(c, stack, axis_to_type<axis_following>());
+                                        return step_do(c, stack, eval, axis_to_type<axis_following>());
                                 
                                 case axis_following_sibling:
-                                        return step_do(c, stack, axis_to_type<axis_following_sibling>());
+                                        return step_do(c, stack, eval, axis_to_type<axis_following_sibling>());
                                 
                                 case axis_namespace:
@@ -8623 +10562 @@
                                 
                                 case axis_parent:
-                                        return step_do(c, stack, axis_to_type<axis_parent>());
+                                        return step_do(c, stack, eval, axis_to_type<axis_parent>());
                                 
                                 case axis_preceding:
-                                        return step_do(c, stack, axis_to_type<axis_preceding>());
+                                        return step_do(c, stack, eval, axis_to_type<axis_preceding>());
 
                                 case axis_preceding_sibling:
-                                        return step_do(c, stack, axis_to_type<axis_preceding_sibling>());
+                                        return step_do(c, stack, eval, axis_to_type<axis_preceding_sibling>());
                                 
                                 case axis_self:
-                                        return step_do(c, stack, axis_to_type<axis_self>());
+                                        return step_do(c, stack, eval, axis_to_type<axis_self>());
 
                                 default:
@@ -8678 +10617 @@
                         }
                 }
+
+                void optimize(xpath_allocator* alloc)
+                {
+                        if (_left) _left->optimize(alloc);
+                        if (_right) _right->optimize(alloc);
+                        if (_next) _next->optimize(alloc);
+
+                        optimize_self(alloc);
+                }
+
+                void optimize_self(xpath_allocator* alloc)
+                {
+                        // Rewrite [position()=expr] with [expr]
+                        // Note that this step has to go before classification to recognize [position()=1]
+                        if ((_type == ast_filter || _type == ast_predicate) &&
+                                _right->_type == ast_op_equal && _right->_left->_type == ast_func_position && _right->_right->_rettype == xpath_type_number)
+                        {
+                                _right = _right->_right;
+                        }
+
+                        // Classify filter/predicate ops to perform various optimizations during evaluation
+                        if (_type == ast_filter || _type == ast_predicate)
+                        {
+                                assert(_test == predicate_default);
+
+                                if (_right->_type == ast_number_constant && _right->_data.number == 1.0)
+                                        _test = predicate_constant_one;
+                                else if (_right->_rettype == xpath_type_number && (_right->_type == ast_number_constant || _right->_type == ast_variable || _right->_type == ast_func_last))
+                                        _test = predicate_constant;
+                                else if (_right->_rettype != xpath_type_number && _right->is_posinv_expr())
+                                        _test = predicate_posinv;
+                        }
+
+                        // Rewrite descendant-or-self::node()/child::foo with descendant::foo
+                        // The former is a full form of //foo, the latter is much faster since it executes the node test immediately
+                        // Do a similar kind of rewrite for self/descendant/descendant-or-self axes
+                        // Note that we only rewrite positionally invariant steps (//foo[1] != /descendant::foo[1])
+                        if (_type == ast_step && (_axis == axis_child || _axis == axis_self || _axis == axis_descendant || _axis == axis_descendant_or_self) && _left &&
+                                _left->_type == ast_step && _left->_axis == axis_descendant_or_self && _left->_test == nodetest_type_node && !_left->_right &&
+                                is_posinv_step())
+                        {
+                                if (_axis == axis_child || _axis == axis_descendant)
+                                        _axis = axis_descendant;
+                                else
+                                        _axis = axis_descendant_or_self;
+
+                                _left = _left->_left;
+                        }
+
+                        // Use optimized lookup table implementation for translate() with constant arguments
+                        if (_type == ast_func_translate && _right->_type == ast_string_constant && _right->_next->_type == ast_string_constant)
+                        {
+                                unsigned char* table = translate_table_generate(alloc, _right->_data.string, _right->_next->_data.string);
+
+                                if (table)
+                                {
+                                        _type = ast_opt_translate_table;
+                                        _data.table = table;
+                                }
+                        }
+
+                        // Use optimized path for @attr = 'value' or @attr = $value
+                        if (_type == ast_op_equal &&
+                                _left->_type == ast_step && _left->_axis == axis_attribute && _left->_test == nodetest_name && !_left->_left && !_left->_right &&
+                                (_right->_type == ast_string_constant || (_right->_type == ast_variable && _right->_rettype == xpath_type_string)))
+                        {
+                                _type = ast_opt_compare_attribute;
+                        }
+                }
                 
-                bool is_posinv()
+                bool is_posinv_expr() const
                 {
                         switch (_type)
                         {
                         case ast_func_position:
+                        case ast_func_last:
                                 return false;
 
@@ -8697 +10706 @@
                         case ast_predicate:
                         case ast_filter:
-                        case ast_filter_posinv:
                                 return true;
 
                         default:
-                                if (_left && !_left->is_posinv()) return false;
+                                if (_left && !_left->is_posinv_expr()) return false;
                                 
                                 for (xpath_ast_node* n = _right; n; n = n->_next)
-                                        if (!n->is_posinv()) return false;
+                                        if (!n->is_posinv_expr()) return false;
                                         
                                 return true;
@@ -8710 +10718 @@
                 }
 
+                bool is_posinv_step() const
+                {
+                        assert(_type == ast_step);
+
+                        for (xpath_ast_node* n = _right; n; n = n->_next)
+                        {
+                                assert(n->_type == ast_predicate);
+
+                                if (n->_test != predicate_posinv)
+                                        return false;
+                        }
+
+                        return true;
+                }
+
                 xpath_value_type rettype() const
                 {
@@ -8725 +10748 @@
 
                 xpath_parse_result* _result;
+
+                char_t _scratch[32];
 
         #ifdef PUGIXML_NO_EXCEPTIONS
@@ -8768 +10793 @@
                                 char_t* c = static_cast<char_t*>(_alloc->allocate_nothrow((length + 1) * sizeof(char_t)));
                                 if (!c) throw_error_oom();
+                                assert(c); // workaround for clang static analysis
 
                                 memcpy(c, value.begin, length * sizeof(char_t));
@@ -8803 +10829 @@
                                 }
                                 else if (name == PUGIXML_TEXT("contains") && argc == 2)
-                                        return new (alloc_node()) xpath_ast_node(ast_func_contains, xpath_type_string, args[0], args[1]);
+                                        return new (alloc_node()) xpath_ast_node(ast_func_contains, xpath_type_boolean, args[0], args[1]);
                                 else if (name == PUGIXML_TEXT("concat") && argc >= 2)
                                         return new (alloc_node()) xpath_ast_node(ast_func_concat, xpath_type_string, args[0], args[1]);
@@ -8865 +10891 @@
                                         return new (alloc_node()) xpath_ast_node(argc == 0 ? ast_func_string_0 : ast_func_string_1, xpath_type_string, args[0]);
                                 else if (name == PUGIXML_TEXT("string-length") && argc <= 1)
-                                        return new (alloc_node()) xpath_ast_node(argc == 0 ? ast_func_string_length_0 : ast_func_string_length_1, xpath_type_string, args[0]);
+                                        return new (alloc_node()) xpath_ast_node(argc == 0 ? ast_func_string_length_0 : ast_func_string_length_1, xpath_type_number, args[0]);
                                 else if (name == PUGIXML_TEXT("starts-with") && argc == 2)
                                         return new (alloc_node()) xpath_ast_node(ast_func_starts_with, xpath_type_boolean, args[0], args[1]);
@@ -9014 +11040 @@
                                         throw_error("Unknown variable: variable set is not provided");
 
-                                xpath_variable* var = get_variable(_variables, name.begin, name.end);
+                                xpath_variable* var = 0;
+                                if (!get_variable_scratch(_scratch, _variables, name.begin, name.end, &var))
+                                        throw_error_oom();
 
                                 if (!var)
@@ -9052 +11080 @@
                                 double value = 0;
 
-                                if (!convert_string_to_number(_lexer.contents().begin, _lexer.contents().end, &value))
+                                if (!convert_string_to_number_scratch(_scratch, _lexer.contents().begin, _lexer.contents().end, &value))
                                         throw_error_oom();
 
@@ -9120 +11148 @@
                                 if (n->rettype() != xpath_type_node_set) throw_error("Predicate has to be applied to node set");
 
-                                bool posinv = expr->rettype() != xpath_type_number && expr->is_posinv();
-
-                                n = new (alloc_node()) xpath_ast_node(posinv ? ast_filter_posinv : ast_filter, xpath_type_node_set, n, expr);
+                                n = new (alloc_node()) xpath_ast_node(ast_filter, n, expr, predicate_default);
 
                                 if (_lexer.current() != lex_close_square_brace)
@@ -9266 +11292 @@
                                 xpath_ast_node* expr = parse_expression();
 
-                                xpath_ast_node* pred = new (alloc_node()) xpath_ast_node(ast_predicate, xpath_type_node_set, expr);
+                                xpath_ast_node* pred = new (alloc_node()) xpath_ast_node(ast_predicate, 0, expr, predicate_default);
                                 
                                 if (_lexer.current() != lex_close_square_brace)
@@ -9277 +11303 @@
                                 last = pred;
                         }
-                        
+
                         return n;
                 }
@@ -9336 +11362 @@
                 //                              | FilterExpr '/' RelativeLocationPath
                 //                              | FilterExpr '//' RelativeLocationPath
-                xpath_ast_node* parse_path_expression()
+                // UnionExpr ::= PathExpr | UnionExpr '|' PathExpr
+                // UnaryExpr ::= UnionExpr | '-' UnaryExpr
+                xpath_ast_node* parse_path_or_unary_expression()
                 {
                         // Clarification.
@@ -9382 +11410 @@
                                 return n;
                         }
-                        else return parse_location_path();
-                }
-
-                // UnionExpr ::= PathExpr | UnionExpr '|' PathExpr
-                xpath_ast_node* parse_union_expression()
-                {
-                        xpath_ast_node* n = parse_path_expression();
-
-                        while (_lexer.current() == lex_union)
+                        else if (_lexer.current() == lex_minus)
                         {
                                 _lexer.next();
 
-                                xpath_ast_node* expr = parse_union_expression();
-
-                                if (n->rettype() != xpath_type_node_set || expr->rettype() != xpath_type_node_set)
+                                // precedence 7+ - only parses union expressions
+                                xpath_ast_node* expr = parse_expression_rec(parse_path_or_unary_expression(), 7);
+
+                                return new (alloc_node()) xpath_ast_node(ast_op_negate, xpath_type_number, expr);
+                        }
+                        else
+                                return parse_location_path();
+                }
+
+                struct binary_op_t
+                {
+                        ast_type_t asttype;
+                        xpath_value_type rettype;
+                        int precedence;
+
+                        binary_op_t(): asttype(ast_unknown), rettype(xpath_type_none), precedence(0)
+                        {
+                        }
+
+                        binary_op_t(ast_type_t asttype_, xpath_value_type rettype_, int precedence_): asttype(asttype_), rettype(rettype_), precedence(precedence_)
+                        {
+                        }
+
+                        static binary_op_t parse(xpath_lexer& lexer)
+                        {
+                                switch (lexer.current())
+                                {
+                                case lex_string:
+                                        if (lexer.contents() == PUGIXML_TEXT("or"))
+                                                return binary_op_t(ast_op_or, xpath_type_boolean, 1);
+                                        else if (lexer.contents() == PUGIXML_TEXT("and"))
+                                                return binary_op_t(ast_op_and, xpath_type_boolean, 2);
+                                        else if (lexer.contents() == PUGIXML_TEXT("div"))
+                                                return binary_op_t(ast_op_divide, xpath_type_number, 6);
+                                        else if (lexer.contents() == PUGIXML_TEXT("mod"))
+                                                return binary_op_t(ast_op_mod, xpath_type_number, 6);
+                                        else
+                                                return binary_op_t();
+
+                                case lex_equal:
+                                        return binary_op_t(ast_op_equal, xpath_type_boolean, 3);
+
+                                case lex_not_equal:
+                                        return binary_op_t(ast_op_not_equal, xpath_type_boolean, 3);
+
+                                case lex_less:
+                                        return binary_op_t(ast_op_less, xpath_type_boolean, 4);
+
+                                case lex_greater:
+                                        return binary_op_t(ast_op_greater, xpath_type_boolean, 4);
+
+                                case lex_less_or_equal:
+                                        return binary_op_t(ast_op_less_or_equal, xpath_type_boolean, 4);
+
+                                case lex_greater_or_equal:
+                                        return binary_op_t(ast_op_greater_or_equal, xpath_type_boolean, 4);
+
+                                case lex_plus:
+                                        return binary_op_t(ast_op_add, xpath_type_number, 5);
+
+                                case lex_minus:
+                                        return binary_op_t(ast_op_subtract, xpath_type_number, 5);
+
+                                case lex_multiply:
+                                        return binary_op_t(ast_op_multiply, xpath_type_number, 6);
+
+                                case lex_union:
+                                        return binary_op_t(ast_op_union, xpath_type_node_set, 7);
+
+                                default:
+                                        return binary_op_t();
+                                }
+                        }
+                };
+
+                xpath_ast_node* parse_expression_rec(xpath_ast_node* lhs, int limit)
+                {
+                        binary_op_t op = binary_op_t::parse(_lexer);
+
+                        while (op.asttype != ast_unknown && op.precedence >= limit)
+                        {
+                                _lexer.next();
+
+                                xpath_ast_node* rhs = parse_path_or_unary_expression();
+
+                                binary_op_t nextop = binary_op_t::parse(_lexer);
+
+                                while (nextop.asttype != ast_unknown && nextop.precedence > op.precedence)
+                                {
+                                        rhs = parse_expression_rec(rhs, nextop.precedence);
+
+                                        nextop = binary_op_t::parse(_lexer);
+                                }
+
+                                if (op.asttype == ast_op_union && (lhs->rettype() != xpath_type_node_set || rhs->rettype() != xpath_type_node_set))
                                         throw_error("Union operator has to be applied to node sets");
 
-                                n = new (alloc_node()) xpath_ast_node(ast_op_union, xpath_type_node_set, n, expr);
-                        }
-
-                        return n;
-                }
-
-                // UnaryExpr ::= UnionExpr | '-' UnaryExpr
-                xpath_ast_node* parse_unary_expression()
-                {
-                        if (_lexer.current() == lex_minus)
-                        {
-                                _lexer.next();
-
-                                xpath_ast_node* expr = parse_unary_expression();
-
-                                return new (alloc_node()) xpath_ast_node(ast_op_negate, xpath_type_number, expr);
-                        }
-                        else return parse_union_expression();
-                }
-                
-                // MultiplicativeExpr ::= UnaryExpr
-                //                                                | MultiplicativeExpr '*' UnaryExpr
-                //                                                | MultiplicativeExpr 'div' UnaryExpr
-                //                                                | MultiplicativeExpr 'mod' UnaryExpr
-                xpath_ast_node* parse_multiplicative_expression()
-                {
-                        xpath_ast_node* n = parse_unary_expression();
-
-                        while (_lexer.current() == lex_multiply || (_lexer.current() == lex_string &&
-                                   (_lexer.contents() == PUGIXML_TEXT("mod") || _lexer.contents() == PUGIXML_TEXT("div"))))
-                        {
-                                ast_type_t op = _lexer.current() == lex_multiply ? ast_op_multiply :
-                                        _lexer.contents().begin[0] == 'd' ? ast_op_divide : ast_op_mod;
-                                _lexer.next();
-
-                                xpath_ast_node* expr = parse_unary_expression();
-
-                                n = new (alloc_node()) xpath_ast_node(op, xpath_type_number, n, expr);
-                        }
-
-                        return n;
-                }
-
-                // AdditiveExpr ::= MultiplicativeExpr
-                //                                      | AdditiveExpr '+' MultiplicativeExpr
-                //                                      | AdditiveExpr '-' MultiplicativeExpr
-                xpath_ast_node* parse_additive_expression()
-                {
-                        xpath_ast_node* n = parse_multiplicative_expression();
-
-                        while (_lexer.current() == lex_plus || _lexer.current() == lex_minus)
-                        {
-                                lexeme_t l = _lexer.current();
-
-                                _lexer.next();
-
-                                xpath_ast_node* expr = parse_multiplicative_expression();
-
-                                n = new (alloc_node()) xpath_ast_node(l == lex_plus ? ast_op_add : ast_op_subtract, xpath_type_number, n, expr);
-                        }
-
-                        return n;
-                }
-
+                                lhs = new (alloc_node()) xpath_ast_node(op.asttype, op.rettype, lhs, rhs);
+
+                                op = binary_op_t::parse(_lexer);
+                        }
+
+                        return lhs;
+                }
+
+                // Expr ::= OrExpr
+                // OrExpr ::= AndExpr | OrExpr 'or' AndExpr
+                // AndExpr ::= EqualityExpr | AndExpr 'and' EqualityExpr
+                // EqualityExpr ::= RelationalExpr
+                //                                      | EqualityExpr '=' RelationalExpr
+                //                                      | EqualityExpr '!=' RelationalExpr
                 // RelationalExpr ::= AdditiveExpr
                 //                                        | RelationalExpr '<' AdditiveExpr
@@ -9468 +11530 @@
                 //                                        | RelationalExpr '<=' AdditiveExpr
                 //                                        | RelationalExpr '>=' AdditiveExpr
-                xpath_ast_node* parse_relational_expression()
-                {
-                        xpath_ast_node* n = parse_additive_expression();
-
-                        while (_lexer.current() == lex_less || _lexer.current() == lex_less_or_equal || 
-                                   _lexer.current() == lex_greater || _lexer.current() == lex_greater_or_equal)
-                        {
-                                lexeme_t l = _lexer.current();
-                                _lexer.next();
-
-                                xpath_ast_node* expr = parse_additive_expression();
-
-                                n = new (alloc_node()) xpath_ast_node(l == lex_less ? ast_op_less : l == lex_greater ? ast_op_greater :
-                                                                l == lex_less_or_equal ? ast_op_less_or_equal : ast_op_greater_or_equal, xpath_type_boolean, n, expr);
-                        }
-
-                        return n;
-                }
-                
-                // EqualityExpr ::= RelationalExpr
-                //                                      | EqualityExpr '=' RelationalExpr
-                //                                      | EqualityExpr '!=' RelationalExpr
-                xpath_ast_node* parse_equality_expression()
-                {
-                        xpath_ast_node* n = parse_relational_expression();
-
-                        while (_lexer.current() == lex_equal || _lexer.current() == lex_not_equal)
-                        {
-                                lexeme_t l = _lexer.current();
-
-                                _lexer.next();
-
-                                xpath_ast_node* expr = parse_relational_expression();
-
-                                n = new (alloc_node()) xpath_ast_node(l == lex_equal ? ast_op_equal : ast_op_not_equal, xpath_type_boolean, n, expr);
-                        }
-
-                        return n;
-                }
-                
-                // AndExpr ::= EqualityExpr | AndExpr 'and' EqualityExpr
-                xpath_ast_node* parse_and_expression()
-                {
-                        xpath_ast_node* n = parse_equality_expression();
-
-                        while (_lexer.current() == lex_string && _lexer.contents() == PUGIXML_TEXT("and"))
-                        {
-                                _lexer.next();
-
-                                xpath_ast_node* expr = parse_equality_expression();
-
-                                n = new (alloc_node()) xpath_ast_node(ast_op_and, xpath_type_boolean, n, expr);
-                        }
-
-                        return n;
-                }
-
-                // OrExpr ::= AndExpr | OrExpr 'or' AndExpr
-                xpath_ast_node* parse_or_expression()
-                {
-                        xpath_ast_node* n = parse_and_expression();
-
-                        while (_lexer.current() == lex_string && _lexer.contents() == PUGIXML_TEXT("or"))
-                        {
-                                _lexer.next();
-
-                                xpath_ast_node* expr = parse_and_expression();
-
-                                n = new (alloc_node()) xpath_ast_node(ast_op_or, xpath_type_boolean, n, expr);
-                        }
-
-                        return n;
-                }
-                
-                // Expr ::= OrExpr
+                // AdditiveExpr ::= MultiplicativeExpr
+                //                                      | AdditiveExpr '+' MultiplicativeExpr
+                //                                      | AdditiveExpr '-' MultiplicativeExpr
+                // MultiplicativeExpr ::= UnaryExpr
+                //                                                | MultiplicativeExpr '*' UnaryExpr
+                //                                                | MultiplicativeExpr 'div' UnaryExpr
+                //                                                | MultiplicativeExpr 'mod' UnaryExpr
                 xpath_ast_node* parse_expression()
                 {
-                        return parse_or_expression();
+                        return parse_expression_rec(parse_path_or_unary_expression(), 0);
                 }
 
@@ -9584 +11578 @@
                 {
                         void* memory = xml_memory::allocate(sizeof(xpath_query_impl));
+                        if (!memory) return 0;
 
                         return new (memory) xpath_query_impl();
                 }
 
-                static void destroy(void* ptr)
-                {
-                        if (!ptr) return;
-                        
+                static void destroy(xpath_query_impl* impl)
+                {
                         // free all allocated pages
-                        static_cast<xpath_query_impl*>(ptr)->alloc.release();
+                        impl->alloc.release();
 
                         // free allocator memory (with the first page)
-                        xml_memory::deallocate(ptr);
+                        xml_memory::deallocate(impl);
                 }
 
@@ -9602 +11595 @@
                 {
                         block.next = 0;
+                        block.capacity = sizeof(block.data);
                 }
 
@@ -9621 +11615 @@
                 return impl->root->eval_string(c, sd.stack);
         }
+
+        PUGI__FN impl::xpath_ast_node* evaluate_node_set_prepare(xpath_query_impl* impl)
+        {
+                if (!impl) return 0;
+
+                if (impl->root->rettype() != xpath_type_node_set)
+                {
+                #ifdef PUGIXML_NO_EXCEPTIONS
+                        return 0;
+                #else
+                        xpath_parse_result res;
+                        res.error = "Expression does not evaluate to node set";
+
+                        throw xpath_exception(res);
+                #endif
+                }
+
+                return impl->root;
+        }
 PUGI__NS_END
 
@@ -9705 +11718 @@
 #endif
 
-        PUGI__FN void xpath_node_set::_assign(const_iterator begin_, const_iterator end_)
+        PUGI__FN void xpath_node_set::_assign(const_iterator begin_, const_iterator end_, type_t type_)
         {
                 assert(begin_ <= end_);
@@ -9721 +11734 @@
                         _begin = &_storage;
                         _end = &_storage + size_;
+                        _type = type_;
                 }
                 else
@@ -9744 +11758 @@
                         _begin = storage;
                         _end = storage + size_;
-                }
-        }
+                        _type = type_;
+                }
+        }
+
+#if __cplusplus >= 201103
+        PUGI__FN void xpath_node_set::_move(xpath_node_set& rhs)
+        {
+                _type = rhs._type;
+                _storage = rhs._storage;
+                _begin = (rhs._begin == &rhs._storage) ? &_storage : rhs._begin;
+                _end = _begin + (rhs._end - rhs._begin);
+
+                rhs._type = type_unsorted;
+                rhs._begin = &rhs._storage;
+                rhs._end = rhs._begin;
+        }
+#endif
 
         PUGI__FN xpath_node_set::xpath_node_set(): _type(type_unsorted), _begin(&_storage), _end(&_storage)
@@ -9751 +11780 @@
         }
 
-        PUGI__FN xpath_node_set::xpath_node_set(const_iterator begin_, const_iterator end_, type_t type_): _type(type_), _begin(&_storage), _end(&_storage)
-        {
-                _assign(begin_, end_);
+        PUGI__FN xpath_node_set::xpath_node_set(const_iterator begin_, const_iterator end_, type_t type_): _type(type_unsorted), _begin(&_storage), _end(&_storage)
+        {
+                _assign(begin_, end_, type_);
         }
 
         PUGI__FN xpath_node_set::~xpath_node_set()
         {
-                if (_begin != &_storage) impl::xml_memory::deallocate(_begin);
+                if (_begin != &_storage)
+                        impl::xml_memory::deallocate(_begin);
         }
                 
-        PUGI__FN xpath_node_set::xpath_node_set(const xpath_node_set& ns): _type(ns._type), _begin(&_storage), _end(&_storage)
-        {
-                _assign(ns._begin, ns._end);
+        PUGI__FN xpath_node_set::xpath_node_set(const xpath_node_set& ns): _type(type_unsorted), _begin(&_storage), _end(&_storage)
+        {
+                _assign(ns._begin, ns._end, ns._type);
         }
         
@@ -9769 +11799 @@
         {
                 if (this == &ns) return *this;
-                
-                _type = ns._type;
-                _assign(ns._begin, ns._end);
+
+                _assign(ns._begin, ns._end, ns._type);
 
                 return *this;
         }
+
+#if __cplusplus >= 201103
+        PUGI__FN xpath_node_set::xpath_node_set(xpath_node_set&& rhs): _type(type_unsorted), _begin(&_storage), _end(&_storage)
+        {
+                _move(rhs);
+        }
+
+        PUGI__FN xpath_node_set& xpath_node_set::operator=(xpath_node_set&& rhs)
+        {
+                if (this == &rhs) return *this;
+
+                if (_begin != &_storage)
+                        impl::xml_memory::deallocate(_begin);
+
+                _move(rhs);
+
+                return *this;
+        }
+#endif
 
         PUGI__FN xpath_node_set::type_t xpath_node_set::type() const
@@ -9831 +11879 @@
         }
 
-        PUGI__FN xpath_variable::xpath_variable()
+        PUGI__FN xpath_variable::xpath_variable(xpath_value_type type_): _type(type_), _next(0)
         {
         }
@@ -9930 +11978 @@
         PUGI__FN xpath_variable_set::xpath_variable_set()
         {
-                for (size_t i = 0; i < sizeof(_data) / sizeof(_data[0]); ++i) _data[i] = 0;
+                for (size_t i = 0; i < sizeof(_data) / sizeof(_data[0]); ++i)
+                        _data[i] = 0;
         }
 
@@ -9936 +11985 @@
         {
                 for (size_t i = 0; i < sizeof(_data) / sizeof(_data[0]); ++i)
-                {
-                        xpath_variable* var = _data[i];
-
-                        while (var)
-                        {
-                                xpath_variable* next = var->_next;
-
-                                impl::delete_xpath_variable(var->_type, var);
-
-                                var = next;
-                        }
-                }
-        }
-
-        PUGI__FN xpath_variable* xpath_variable_set::find(const char_t* name) const
+                        _destroy(_data[i]);
+        }
+
+        PUGI__FN xpath_variable_set::xpath_variable_set(const xpath_variable_set& rhs)
+        {
+                for (size_t i = 0; i < sizeof(_data) / sizeof(_data[0]); ++i)
+                        _data[i] = 0;
+
+                _assign(rhs);
+        }
+
+        PUGI__FN xpath_variable_set& xpath_variable_set::operator=(const xpath_variable_set& rhs)
+        {
+                if (this == &rhs) return *this;
+
+                _assign(rhs);
+
+                return *this;
+        }
+
+#if __cplusplus >= 201103
+        PUGI__FN xpath_variable_set::xpath_variable_set(xpath_variable_set&& rhs)
+        {
+                for (size_t i = 0; i < sizeof(_data) / sizeof(_data[0]); ++i)
+                {
+                        _data[i] = rhs._data[i];
+                        rhs._data[i] = 0;
+                }
+        }
+
+        PUGI__FN xpath_variable_set& xpath_variable_set::operator=(xpath_variable_set&& rhs)
+        {
+                for (size_t i = 0; i < sizeof(_data) / sizeof(_data[0]); ++i)
+                {
+                        _destroy(_data[i]);
+
+                        _data[i] = rhs._data[i];
+                        rhs._data[i] = 0;
+                }
+
+                return *this;
+        }
+#endif
+
+        PUGI__FN void xpath_variable_set::_assign(const xpath_variable_set& rhs)
+        {
+                xpath_variable_set temp;
+
+                for (size_t i = 0; i < sizeof(_data) / sizeof(_data[0]); ++i)
+                        if (rhs._data[i] && !_clone(rhs._data[i], &temp._data[i]))
+                                return;
+
+                _swap(temp);
+        }
+
+        PUGI__FN void xpath_variable_set::_swap(xpath_variable_set& rhs)
+        {
+                for (size_t i = 0; i < sizeof(_data) / sizeof(_data[0]); ++i)
+                {
+                        xpath_variable* chain = _data[i];
+
+                        _data[i] = rhs._data[i];
+                        rhs._data[i] = chain;
+                }
+        }
+
+        PUGI__FN xpath_variable* xpath_variable_set::_find(const char_t* name) const
         {
                 const size_t hash_size = sizeof(_data) / sizeof(_data[0]);
@@ -9963 +12064 @@
         }
 
+        PUGI__FN bool xpath_variable_set::_clone(xpath_variable* var, xpath_variable** out_result)
+        {
+                xpath_variable* last = 0;
+
+                while (var)
+                {
+                        // allocate storage for new variable
+                        xpath_variable* nvar = impl::new_xpath_variable(var->_type, var->name());
+                        if (!nvar) return false;
+
+                        // link the variable to the result immediately to handle failures gracefully
+                        if (last)
+                                last->_next = nvar;
+                        else
+                                *out_result = nvar;
+
+                        last = nvar;
+
+                        // copy the value; this can fail due to out-of-memory conditions
+                        if (!impl::copy_xpath_variable(nvar, var)) return false;
+
+                        var = var->_next;
+                }
+
+                return true;
+        }
+
+        PUGI__FN void xpath_variable_set::_destroy(xpath_variable* var)
+        {
+                while (var)
+                {
+                        xpath_variable* next = var->_next;
+
+                        impl::delete_xpath_variable(var->_type, var);
+
+                        var = next;
+                }
+        }
+
         PUGI__FN xpath_variable* xpath_variable_set::add(const char_t* name, xpath_value_type type)
         {
@@ -9978 +12118 @@
                 if (result)
                 {
-                        result->_type = type;
                         result->_next = _data[hash];
 
@@ -10013 +12152 @@
         PUGI__FN xpath_variable* xpath_variable_set::get(const char_t* name)
         {
-                return find(name);
+                return _find(name);
         }
 
         PUGI__FN const xpath_variable* xpath_variable_set::get(const char_t* name) const
         {
-                return find(name);
+                return _find(name);
         }
 
@@ -10035 +12174 @@
                 else
                 {
-                        impl::buffer_holder impl_holder(qimpl, impl::xpath_query_impl::destroy);
+                        using impl::auto_deleter; // MSVC7 workaround
+                        auto_deleter<impl::xpath_query_impl> impl(qimpl, impl::xpath_query_impl::destroy);
 
                         qimpl->root = impl::xpath_parser::parse(query, variables, &qimpl->alloc, &_result);
@@ -10041 +12181 @@
                         if (qimpl->root)
                         {
-                                _impl = static_cast<impl::xpath_query_impl*>(impl_holder.release());
+                                qimpl->root->optimize(&qimpl->alloc);
+
+                                _impl = impl.release();
                                 _result.error = 0;
                         }
@@ -10047 +12189 @@
         }
 
+        PUGI__FN xpath_query::xpath_query(): _impl(0)
+        {
+        }
+
         PUGI__FN xpath_query::~xpath_query()
         {
-                impl::xpath_query_impl::destroy(_impl);
-        }
+                if (_impl)
+                        impl::xpath_query_impl::destroy(static_cast<impl::xpath_query_impl*>(_impl));
+        }
+
+#if __cplusplus >= 201103
+        PUGI__FN xpath_query::xpath_query(xpath_query&& rhs)
+        {
+                _impl = rhs._impl;
+                _result = rhs._result;
+                rhs._impl = 0;
+                rhs._result = xpath_parse_result();
+        }
+
+        PUGI__FN xpath_query& xpath_query::operator=(xpath_query&& rhs)
+        {
+                if (this == &rhs) return *this;
+
+                if (_impl)
+                        impl::xpath_query_impl::destroy(static_cast<impl::xpath_query_impl*>(_impl));
+
+                _impl = rhs._impl;
+                _result = rhs._result;
+                rhs._impl = 0;
+                rhs._result = xpath_parse_result();
+
+                return *this;
+        }
+#endif
 
         PUGI__FN xpath_value_type xpath_query::return_type() const
@@ -10092 +12264 @@
                 impl::xpath_stack_data sd;
 
-                return impl::evaluate_string_impl(static_cast<impl::xpath_query_impl*>(_impl), n, sd).c_str();
+                impl::xpath_string r = impl::evaluate_string_impl(static_cast<impl::xpath_query_impl*>(_impl), n, sd);
+
+                return string_t(r.c_str(), r.length());
         }
 #endif
@@ -10118 +12292 @@
         PUGI__FN xpath_node_set xpath_query::evaluate_node_set(const xpath_node& n) const
         {
-                if (!_impl) return xpath_node_set();
-
-                impl::xpath_ast_node* root = static_cast<impl::xpath_query_impl*>(_impl)->root;
-
-                if (root->rettype() != xpath_type_node_set)
-                {
-                #ifdef PUGIXML_NO_EXCEPTIONS
-                        return xpath_node_set();
-                #else
-                        xpath_parse_result res;
-                        res.error = "Expression does not evaluate to node set";
-
-                        throw xpath_exception(res);
-                #endif
-                }
-                
+                impl::xpath_ast_node* root = impl::evaluate_node_set_prepare(static_cast<impl::xpath_query_impl*>(_impl));
+                if (!root) return xpath_node_set();
+
                 impl::xpath_context c(n, 1, 1);
                 impl::xpath_stack_data sd;
@@ -10141 +12302 @@
         #endif
 
-                impl::xpath_node_set_raw r = root->eval_node_set(c, sd.stack);
+                impl::xpath_node_set_raw r = root->eval_node_set(c, sd.stack, impl::nodeset_eval_all);
 
                 return xpath_node_set(r.begin(), r.end(), r.type());
         }
 
+        PUGI__FN xpath_node xpath_query::evaluate_node(const xpath_node& n) const
+        {
+                impl::xpath_ast_node* root = impl::evaluate_node_set_prepare(static_cast<impl::xpath_query_impl*>(_impl));
+                if (!root) return xpath_node();
+
+                impl::xpath_context c(n, 1, 1);
+                impl::xpath_stack_data sd;
+
+        #ifdef PUGIXML_NO_EXCEPTIONS
+                if (setjmp(sd.error_handler)) return xpath_node();
+        #endif
+
+                impl::xpath_node_set_raw r = root->eval_node_set(c, sd.stack, impl::nodeset_eval_first);
+
+                return r.first();
+        }
+
         PUGI__FN const xpath_parse_result& xpath_query::result() const
         {
@@ -10163 +12341 @@
         {
                 return !_impl;
+        }
+
+        PUGI__FN xpath_node xml_node::select_node(const char_t* query, xpath_variable_set* variables) const
+        {
+                xpath_query q(query, variables);
+                return select_node(q);
+        }
+
+        PUGI__FN xpath_node xml_node::select_node(const xpath_query& query) const
+        {
+                return query.evaluate_node(*this);
+        }
+
+        PUGI__FN xpath_node_set xml_node::select_nodes(const char_t* query, xpath_variable_set* variables) const
+        {
+                xpath_query q(query, variables);
+                return select_nodes(q);
+        }
+
+        PUGI__FN xpath_node_set xml_node::select_nodes(const xpath_query& query) const
+        {
+                return query.evaluate_node_set(*this);
         }
 
@@ -10173 +12373 @@
         PUGI__FN xpath_node xml_node::select_single_node(const xpath_query& query) const
         {
-                xpath_node_set s = query.evaluate_node_set(*this);
-                return s.empty() ? xpath_node() : s.first();
-        }
-
-        PUGI__FN xpath_node_set xml_node::select_nodes(const char_t* query, xpath_variable_set* variables) const
-        {
-                xpath_query q(query, variables);
-                return select_nodes(q);
-        }
-
-        PUGI__FN xpath_node_set xml_node::select_nodes(const xpath_query& query) const
-        {
-                return query.evaluate_node_set(*this);
+                return query.evaluate_node(*this);
         }
 }
@@ -10203 +12391 @@
 // Undefine all local macros (makes sure we're not leaking macros in header-only mode)
 #undef PUGI__NO_INLINE
+#undef PUGI__UNLIKELY
 #undef PUGI__STATIC_ASSERT
 #undef PUGI__DMC_VOLATILE
@@ -10210 +12399 @@
 #undef PUGI__FN
 #undef PUGI__FN_NO_INLINE
+#undef PUGI__GETPAGE_IMPL
+#undef PUGI__GETPAGE
+#undef PUGI__NODETYPE
 #undef PUGI__IS_CHARTYPE_IMPL
 #undef PUGI__IS_CHARTYPE
 #undef PUGI__IS_CHARTYPEX
+#undef PUGI__ENDSWITH
 #undef PUGI__SKIPWS
 #undef PUGI__OPTSET
@@ -10219 +12412 @@
 #undef PUGI__SCANFOR
 #undef PUGI__SCANWHILE
+#undef PUGI__SCANWHILE_UNROLL
 #undef PUGI__ENDSEG
 #undef PUGI__THROW_ERROR
@@ -10226 +12420 @@
 
 /**
- * Copyright (c) 2006-2012 Arseny Kapoulkine
+ * Copyright (c) 2006-2015 Arseny Kapoulkine
  *
  * Permission is hereby granted, free of charge, to any person