pugixml 1.7 manual

pugixml is a light-weight C++ XML processing library. It consists of a DOM-like interface with rich traversal/modification capabilities, an extremely fast XML parser which constructs the DOM tree from an XML file/buffer, and an XPath 1.0 implementation for complex data-driven tree queries. Full Unicode support is also available, with two Unicode interface variants and conversions between different Unicode encodings (which happen automatically during parsing/saving). The library is extremely portable and easy to integrate and use. pugixml is developed and maintained since 2006 and has many users. All code is distributed under the MIT license, making it completely free to use in both open-source and proprietary applications.

pugixml enables very fast, convenient and memory-efficient XML document processing. However, since pugixml has a DOM parser, it can’t process XML documents that do not fit in memory; also the parser is a non-validating one, so if you need DTD or XML Schema validation, the library is not for you.

This is the complete manual for pugixml, which describes all features of the library in detail. If you want to start writing code as quickly as possible, you are advised to read the quick start guide first.

If you believe you’ve found a bug in pugixml (bugs include compilation problems (errors/warnings), crashes, performance degradation and incorrect behavior), please file an issue via issue submission form. Be sure to include the relevant information so that the bug can be reproduced: the version of pugixml, compiler version and target architecture, the code that uses pugixml and exhibits the bug, etc.

Feature requests can be reported the same way as bugs, so if you’re missing some functionality in pugixml or if the API is rough in some places and you can suggest an improvement, file an issue. However please note that there are many factors when considering API changes (compatibility with previous versions, API redundancy, etc.), so generally features that can be implemented via a small function without pugixml modification are not accepted. However, all rules have exceptions.

If you have a contribution to pugixml, such as build script for some build system/IDE, or a well-designed set of helper functions, or a binding to some language other than C++, please file an issue or open a pull request. Your contribution has to be distributed under the terms of a license that’s compatible with pugixml license; i.e. GPL/LGPL licensed code is not accepted.

If filing an issue is not possible due to privacy or other concerns, you can contact pugixml author by e-mail directly: arseny.kapoulkine@gmail.com.

pugixml could not be developed without the help from many people; some of them are listed in this section. If you’ve played a part in pugixml development and you can not find yourself on this list, I’m truly sorry; please send me an e-mail so I can fix this.

Thanks to Kristen Wegner for pugxml parser, which was used as a basis for pugixml.

Thanks to Neville Franks for contributions to pugxml parser.

Thanks to Artyom Palvelev for suggesting a lazy gap contraction approach.

Thanks to Vyacheslav Egorov for documentation proofreading and fuzz testing.

The pugixml library is distributed under the MIT license:

This means that you can freely use pugixml in your applications, both open-source and proprietary. If you use pugixml in a product, it is sufficient to add an acknowledgment like this to the product distribution:

pugixml is distributed in source form. You can either download a source distribution or clone the Git repository.

pugixml is distributed in source form without any pre-built binaries; you have to build them yourself.

The complete pugixml source consists of three files - one source file, pugixml.cpp, and two header files, pugixml.hpp and pugiconfig.hpp. pugixml.hpp is the primary header which you need to include in order to use pugixml classes/functions; pugiconfig.hpp is a supplementary configuration file (see Additional configuration options). The rest of this guide assumes that pugixml.hpp is either in the current directory or in one of include directories of your projects, so that #include "pugixml.hpp" can find the header; however you can also use relative path (i.e. #include "../libs/pugixml/src/pugixml.hpp") or include directory-relative path (i.e. #include <xml/thirdparty/pugixml/src/pugixml.hpp>).

pugixml is written in standard-compliant C++ with some compiler-specific workarounds where appropriate. pugixml is compatible with the C++11 standard, but does not require C++11 support. Each version is tested with a unit test suite (with code coverage about 99%) on the following platforms:

The XML document is represented with a tree data structure. The root of the tree is the document itself, which corresponds to C++ type xml_document. Document has one or more child nodes, which correspond to C++ type xml_node. Nodes have different types; depending on a type, a node can have a collection of child nodes, a collection of attributes, which correspond to C++ type xml_attribute, and some additional data (i.e. name).

The tree nodes can be of one of the following types (which together form the enumeration xml_node_type):

Finally, here is a complete example of XML document and the corresponding tree representation (samples/tree.xml):

Despite the fact that there are several node types, there are only three C++ classes representing the tree (xml_document, xml_node, xml_attribute); some operations on xml_node are only valid for certain node types. The classes are described below.

xml_document is the owner of the entire document structure; it is a non-copyable class. The interface of xml_document consists of loading functions (see Loading document), saving functions (see Saving document) and the entire interface of xml_node, which allows for document inspection and/or modification. Note that while xml_document is a sub-class of xml_node, xml_node is not a polymorphic type; the inheritance is present only to simplify usage. Alternatively you can use the document_element function to get the element node that’s the immediate child of the document.

Default constructor of xml_document initializes the document to the tree with only a root node (document node). You can then populate it with data using either tree modification functions or loading functions; all loading functions destroy the previous tree with all occupied memory, which puts existing node/attribute handles for this document to invalid state. If you want to destroy the previous tree, you can use the xml_document::reset function; it destroys the tree and replaces it with either an empty one or a copy of the specified document. Destructor of xml_document also destroys the tree, thus the lifetime of the document object should exceed the lifetimes of any node/attribute handles that point to the tree.

xml_node is the handle to document node; it can point to any node in the document, including the document node itself. There is a common interface for nodes of all types; the actual node type can be queried via the xml_node::type() method. Note that xml_node is only a handle to the actual node, not the node itself - you can have several xml_node handles pointing to the same underlying object. Destroying xml_node handle does not destroy the node and does not remove it from the tree. The size of xml_node is equal to that of a pointer, so it is nothing more than a lightweight wrapper around a pointer; you can safely pass or return xml_node objects by value without additional overhead.

There is a special value of xml_node type, known as null node or empty node (such nodes have type node_null). It does not correspond to any node in any document, and thus resembles null pointer. However, all operations are defined on empty nodes; generally the operations don’t do anything and return empty nodes/attributes or empty strings as their result (see documentation for specific functions for more detailed information). This is useful for chaining calls; i.e. you can get the grandparent of a node like so: node.parent().parent(); if a node is a null node or it does not have a parent, the first parent() call returns null node; the second parent() call then also returns null node, which makes error handling easier.

xml_attribute is the handle to an XML attribute; it has the same semantics as xml_node, i.e. there can be several xml_attribute handles pointing to the same underlying object and there is a special null attribute value, which propagates to function results.

Both xml_node and xml_attribute have the default constructor which initializes them to null objects.

xml_node and xml_attribute try to behave like pointers, that is, they can be compared with other objects of the same type, making it possible to use them as keys in associative containers. All handles to the same underlying object are equal, and any two handles to different underlying objects are not equal. Null handles only compare as equal to themselves. The result of relational comparison can not be reliably determined from the order of nodes in file or in any other way. Do not use relational comparison operators except for search optimization (i.e. associative container keys).

If you want to use xml_node or xml_attribute objects as keys in hash-based associative containers, you can use the hash_value member functions. They return the hash values that are guaranteed to be the same for all handles to the same underlying object. The hash value for null handles is 0. Note that hash value does not depend on the content of the node, only on the location of the underlying structure in memory - this means that loading the same document twice will likely produce different hash values, and copying the node will not preserve the hash.

Finally handles can be implicitly cast to boolean-like objects, so that you can test if the node/attribute is empty with the following code: if (node) { …​ } or if (!node) { …​ } else { …​ }. Alternatively you can check if a given xml_node/xml_attribute handle is null by calling the following methods:

Nodes and attributes do not exist without a document tree, so you can’t create them without adding them to some document. Once underlying node/attribute objects are destroyed, the handles to those objects become invalid. While this means that destruction of the entire tree invalidates all node/attribute handles, it also means that destroying a subtree (by calling xml_node::remove_child) or removing an attribute invalidates the corresponding handles. There is no way to check handle validity; you have to ensure correctness through external mechanisms.

There are two choices of interface and internal representation when configuring pugixml: you can either choose the UTF-8 (also called char) interface or UTF-16/32 (also called wchar_t) one. The choice is controlled via PUGIXML_WCHAR_MODE define; you can set it via pugiconfig.hpp or via preprocessor options, as discussed in Additional configuration options. If this define is set, the wchar_t interface is used; otherwise (by default) the char interface is used. The exact wide character encoding is assumed to be either UTF-16 or UTF-32 and is determined based on the size of wchar_t type.

All tree functions that work with strings work with either C-style null terminated strings or STL strings of the selected character type. For example, node name accessors look like this in char mode:

and like this in wchar_t mode:

There is a special type, pugi::char_t, that is defined as the character type and depends on the library configuration; it will be also used in the documentation hereafter. There is also a type pugi::string_t, which is defined as the STL string of the character type; it corresponds to std::string in char mode and to std::wstring in wchar_t mode.

In addition to the interface, the internal implementation changes to store XML data as pugi::char_t; this means that these two modes have different memory usage characteristics - generally UTF-8 mode is more memory and performance efficient, especially if sizeof(wchar_t) is 4. The conversion to pugi::char_t upon document loading and from pugi::char_t upon document saving happen automatically, which also carries minor performance penalty. The general advice however is to select the character mode based on usage scenario, i.e. if UTF-8 is inconvenient to process and most of your XML data is non-ASCII, wchar_t mode is probably a better choice.

There are cases when you’ll have to convert string data between UTF-8 and wchar_t encodings; the following helper functions are provided for such purposes:

Both functions accept a null-terminated string as an argument str, and return the converted string. as_utf8 performs conversion from UTF-16/32 to UTF-8; as_wide performs conversion from UTF-8 to UTF-16/32. Invalid UTF sequences are silently discarded upon conversion. str has to be a valid string; passing null pointer results in undefined behavior. There are also two overloads with the same semantics which accept a string as an argument:

Almost all functions in pugixml have the following thread-safety guarantees:

Concurrent modification and traversing of a single tree requires synchronization, for example via reader-writer lock. Modification includes altering document structure and altering individual node/attribute data, i.e. changing names/values.

The only exception is set_memory_management_functions; it modifies global variables and as such is not thread-safe. Its usage policy has more restrictions, see Custom memory allocation/deallocation functions.

With the exception of XPath, pugixml itself does not throw any exceptions. Additionally, most pugixml functions have a no-throw exception guarantee.

This is not applicable to functions that operate on STL strings or IOstreams; such functions have either strong guarantee (functions that operate on strings) or basic guarantee (functions that operate on streams). Also functions that call user-defined callbacks (i.e. xml_node::traverse or xml_node::find_node) do not provide any exception guarantees beyond the ones provided by the callback.

If exception handling is not disabled with PUGIXML_NO_EXCEPTIONS define, XPath functions may throw xpath_exception on parsing errors; also, XPath functions may throw std::bad_alloc in low memory conditions. Still, XPath functions provide strong exception guarantee.

pugixml requests the memory needed for document storage in big chunks, and allocates document data inside those chunks. This section discusses replacing functions used for chunk allocation and internal memory management implementation.

The most common source of XML data is files; pugixml provides dedicated functions for loading an XML document from file:

These functions accept the file path as its first argument, and also two optional arguments, which specify parsing options (see Parsing options) and input data encoding (see Encodings). The path has the target operating system format, so it can be a relative or absolute one, it should have the delimiters of the target system, it should have the exact case if the target file system is case-sensitive, etc.

File path is passed to the system file opening function as is in case of the first function (which accepts const char* path); the second function either uses a special file opening function if it is provided by the runtime library or converts the path to UTF-8 and uses the system file opening function.

load_file destroys the existing document tree and then tries to load the new tree from the specified file. The result of the operation is returned in an xml_parse_result object; this object contains the operation status and the related information (i.e. last successfully parsed position in the input file, if parsing fails). See Handling parsing errors for error handling details.

This is an example of loading XML document from file (samples/load_file.cpp):

Sometimes XML data should be loaded from some other source than a file, i.e. HTTP URL; also you may want to load XML data from file using non-standard functions, i.e. to use your virtual file system facilities or to load XML from GZip-compressed files. All these scenarios require loading document from memory. First you should prepare a contiguous memory block with all XML data; then you have to invoke one of buffer loading functions. These functions will handle the necessary encoding conversions, if any, and then will parse the data into the corresponding XML tree. There are several buffer loading functions, which differ in the behavior and thus in performance/memory usage:

All functions accept the buffer which is represented by a pointer to XML data, contents, and data size in bytes. Also there are two optional arguments, which specify parsing options (see Parsing options) and input data encoding (see Encodings). The buffer does not have to be zero-terminated.

load_buffer function works with immutable buffer - it does not ever modify the buffer. Because of this restriction it has to create a private buffer and copy XML data to it before parsing (applying encoding conversions if necessary). This copy operation carries a performance penalty, so inplace functions are provided - load_buffer_inplace and load_buffer_inplace_own store the document data in the buffer, modifying it in the process. In order for the document to stay valid, you have to make sure that the buffer’s lifetime exceeds that of the tree if you’re using inplace functions. In addition to that, load_buffer_inplace does not assume ownership of the buffer, so you’ll have to destroy it yourself; load_buffer_inplace_own assumes ownership of the buffer and destroys it once it is not needed. This means that if you’re using load_buffer_inplace_own, you have to allocate memory with pugixml allocation function (you can get it via get_memory_allocation_function).

The best way from the performance/memory point of view is to load document using load_buffer_inplace_own; this function has maximum control of the buffer with XML data so it is able to avoid redundant copies and reduce peak memory usage while parsing. This is the recommended function if you have to load the document from memory and performance is critical.

There is also a simple helper function for cases when you want to load the XML document from null-terminated character string:

It is equivalent to calling load_buffer with size being either strlen(contents) or wcslen(contents) * sizeof(wchar_t), depending on the character type. This function assumes native encoding for input data, so it does not do any encoding conversion. In general, this function is fine for loading small documents from string literals, but has more overhead and less functionality than the buffer loading functions.

This is an example of loading XML document from memory using different functions (samples/load_memory.cpp):

To enhance interoperability, pugixml provides functions for loading document from any object which implements C++ std::istream interface. This allows you to load documents from any standard C++ stream (i.e. file stream) or any third-party compliant implementation (i.e. Boost Iostreams). There are two functions, one works with narrow character streams, another handles wide character ones:

load with std::istream argument loads the document from stream from the current read position to the end, treating the stream contents as a byte stream of the specified encoding (with encoding autodetection as necessary). Thus calling xml_document::load on an opened std::ifstream object is equivalent to calling xml_document::load_file.

load with std::wstream argument treats the stream contents as a wide character stream (encoding is always encoding_wchar). Because of this, using load with wide character streams requires careful (usually platform-specific) stream setup (i.e. using the imbue function). Generally use of wide streams is discouraged, however it provides you the ability to load documents from non-Unicode encodings, i.e. you can load Shift-JIS encoded data if you set the correct locale.

This is a simple example of loading XML document from file using streams (samples/load_stream.cpp); read the sample code for more complex examples involving wide streams and locales:

All document loading functions return the parsing result via xml_parse_result object. It contains parsing status, the offset of last successfully parsed character from the beginning of the source stream, and the encoding of the source stream:

Parsing status is represented as the xml_parse_status enumeration and can be one of the following:

description() member function can be used to convert parsing status to a string; the returned message is always in English, so you’ll have to write your own function if you need a localized string. However please note that the exact messages returned by description() function may change from version to version, so any complex status handling should be based on status value. Note that description() returns a char string even in PUGIXML_WCHAR_MODE; you’ll have to call as_wide to get the wchar_t string.

If parsing failed because the source data was not a valid XML, the resulting tree is not destroyed - despite the fact that load function returns error, you can use the part of the tree that was successfully parsed. Obviously, the last element may have an unexpected name/value; for example, if the attribute value does not end with the necessary quotation mark, like in <node attr="value>some data</node> example, the value of attribute attr will contain the string value>some data</node>.

In addition to the status code, parsing result has an offset member, which contains the offset of last successfully parsed character if parsing failed because of an error in source data; otherwise offset is 0. For parsing efficiency reasons, pugixml does not track the current line during parsing; this offset is in units of pugi::char_t (bytes for character mode, wide characters for wide character mode). Many text editors support 'Go To Position' feature - you can use it to locate the exact error position. Alternatively, if you’re loading the document from memory, you can display the error chunk along with the error description (see the example code below).

Parsing result also has an encoding member, which can be used to check that the source data encoding was correctly guessed. It is equal to the exact encoding used during parsing (i.e. with the exact endianness); see Encodings for more information.

Parsing result object can be implicitly converted to bool; if you do not want to handle parsing errors thoroughly, you can just check the return value of load functions as if it was a bool: if (doc.load_file("file.xml")) { …​ } else { …​ }.

This is an example of handling loading errors (samples/load_error_handling.cpp):

All document loading functions accept the optional parameter options. This is a bitmask that customizes the parsing process: you can select the node types that are parsed and various transformations that are performed with the XML text. Disabling certain transformations can improve parsing performance for some documents; however, the code for all transformations is very well optimized, and thus the majority of documents won’t get any performance benefit. As a rule of thumb, only modify parsing flags if you want to get some nodes in the document that are excluded by default (i.e. declaration or comment nodes).

These flags control the resulting tree contents:

These flags control the transformation of tree element contents:

Additionally there are three predefined option masks:

This is an example of using different parsing options (samples/load_options.cpp):

pugixml supports all popular Unicode encodings (UTF-8, UTF-16 (big and little endian), UTF-32 (big and little endian); UCS-2 is naturally supported since it’s a strict subset of UTF-16) and handles all encoding conversions. Most loading functions accept the optional parameter encoding. This is a value of enumeration type xml_encoding, that can have the following values:

The algorithm used for encoding_auto correctly detects any supported Unicode encoding for all well-formed XML documents (since they start with document declaration) and for all other XML documents that start with <; if your XML document does not start with < and has encoding that is different from UTF-8, use the specific encoding.

pugixml is not fully W3C conformant - it can load any valid XML document, but does not perform some well-formedness checks. While considerable effort is made to reject invalid XML documents, some validation is not performed because of performance reasons.

There is only one non-conformant behavior when dealing with valid XML documents: pugixml does not use information supplied in document type declaration for parsing. This means that entities declared in DOCTYPE are not expanded, and all attribute/PCDATA values are always processed in a uniform way that depends only on parsing options.

As for rejecting invalid XML documents, there are a number of incompatibilities with W3C specification, including:

The internal representation of the document is a tree, where each node has a list of child nodes (the order of children corresponds to their order in the XML representation), and additionally element nodes have a list of attributes, which is also ordered. Several functions are provided in order to let you get from one node in the tree to the other. These functions roughly correspond to the internal representation, and thus are usually building blocks for other methods of traversing (i.e. XPath traversals are based on these functions).

parent function returns the node’s parent; all non-null nodes except the document have non-null parent. first_child and last_child return the first and last child of the node, respectively; note that only document nodes and element nodes can have non-empty child node list. If node has no children, both functions return null nodes. next_sibling and previous_sibling return the node that’s immediately to the right/left of this node in the children list, respectively - for example, in <a/><b/><c/>, calling next_sibling for a handle that points to <b/> results in a handle pointing to <c/>, and calling previous_sibling results in handle pointing to <a/>. If node does not have next/previous sibling (this happens if it is the last/first node in the list, respectively), the functions return null nodes. first_attribute, last_attribute, next_attribute and previous_attribute functions behave similarly to the corresponding child node functions and allow to iterate through attribute list in the same way.

Calling any of the functions above on the null handle results in a null handle - i.e. node.first_child().next_sibling() returns the second child of node, and null handle if node is null, has no children at all or if it has only one child node.

With these functions, you can iterate through all child nodes and display all attributes like this (samples/traverse_base.cpp):

Apart from structural information (parent, child nodes, attributes), nodes can have name and value, both of which are strings. Depending on node type, name or value may be absent. node_document nodes do not have a name or value, node_element and node_declaration nodes always have a name but never have a value, node_pcdata, node_cdata, node_comment and node_doctype nodes never have a name but always have a value (it may be empty though), node_pi nodes always have a name and a value (again, value may be empty). In order to get node’s name or value, you can use the following functions:

In case node does not have a name or value or if the node handle is null, both functions return empty strings - they never return null pointers.

It is common to store data as text contents of some node - i.e. <node><description>This is a node</description></node>. In this case, <description> node does not have a value, but instead has a child of type node_pcdata with value "This is a node". pugixml provides several helper functions to parse such data:

child_value() returns the value of the first child with type node_pcdata or node_cdata; child_value(name) is a simple wrapper for child(name).child_value(). For the above example, calling node.child_value("description") and description.child_value() will both produce string "This is a node". If there is no child with relevant type, or if the handle is null, child_value functions return empty string.

text() returns a special object that can be used for working with PCDATA contents in more complex cases than just retrieving the value; it is described in Working with text contents sections.

There is an example of using some of these functions at the end of the next section.

All attributes have name and value, both of which are strings (value may be empty). There are two corresponding accessors, like for xml_node:

In case the attribute handle is null, both functions return empty strings - they never return null pointers.

If you need a non-empty string if the attribute handle is null (for example, you need to get the option value from XML attribute, but if it is not specified, you need it to default to "sorted" instead of ""), you can use as_string accessor:

It returns def argument if the attribute handle is null. If you do not specify the argument, the function is equivalent to value().

In many cases attribute values have types that are not strings - i.e. an attribute may always contain values that should be treated as integers, despite the fact that they are represented as strings in XML. pugixml provides several accessors that convert attribute value to some other type:

as_int, as_uint, as_llong, as_ullong, as_double and as_float convert attribute values to numbers. If attribute handle is null or attribute value is empty, def argument is returned (which is 0 by default). Otherwise, all leading whitespace characters are truncated, and the remaining string is parsed as an integer number in either decimal or hexadecimal form (applicable to as_int, as_uint, as_llong and as_ullong; hexadecimal format is used if the number has 0x or 0X prefix) or as a floating point number in either decimal or scientific form (as_double or as_float). Any extra characters are silently discarded, i.e. as_int will return 1 for string "1abc".

In case the input string contains a number that is out of the target numeric range, the result is undefined.

as_bool converts attribute value to boolean as follows: if attribute handle is null, def argument is returned (which is false by default). If attribute value is empty, false is returned. Otherwise, true is returned if the first character is one of '1', 't', 'T', 'y', 'Y'. This means that strings like "true" and "yes" are recognized as true, while strings like "false" and "no" are recognized as false. For more complex matching you’ll have to write your own function.

This is an example of using these functions, along with node data retrieval ones (samples/traverse_base.cpp):

Since a lot of document traversal consists of finding the node/attribute with the correct name, there are special functions for that purpose:

child and attribute return the first child/attribute with the specified name; next_sibling and previous_sibling return the first sibling in the corresponding direction with the specified name. All string comparisons are case-sensitive. In case the node handle is null or there is no node/attribute with the specified name, null handle is returned.

child and next_sibling functions can be used together to loop through all child nodes with the desired name like this:

Occasionally the needed node is specified not by the unique name but instead by the value of some attribute; for example, it is common to have node collections with each node having a unique id: <group><item id="1"/> <item id="2"/></group>. There are two functions for finding child nodes based on the attribute values:

The three-argument function returns the first child node with the specified name which has an attribute with the specified name/value; the two-argument function skips the name test for the node, which can be useful for searching in heterogeneous collections. If the node handle is null or if no node is found, null handle is returned. All string comparisons are case-sensitive.

In all of the above functions, all arguments have to be valid strings; passing null pointers results in undefined behavior.

This is an example of using these functions (samples/traverse_base.cpp):

If your C++ compiler supports range-based for-loop (this is a C++11 feature, at the time of writing it’s supported by Microsoft Visual Studio 2012+, GCC 4.6+ and Clang 3.0+), you can use it to enumerate nodes/attributes. Additional helpers are provided to support this; note that they are also compatible with Boost Foreach, and possibly other pre-C++11 foreach facilities.

children function allows you to enumerate all child nodes; children function with name argument allows you to enumerate all child nodes with a specific name; attributes function allows you to enumerate all attributes of the node. Note that you can also use node object itself in a range-based for construct, which is equivalent to using children().

This is an example of using these functions (samples/traverse_rangefor.cpp):

Child node lists and attribute lists are simply double-linked lists; while you can use previous_sibling/next_sibling and other such functions for iteration, pugixml additionally provides node and attribute iterators, so that you can treat nodes as containers of other nodes or attributes:

begin and attributes_begin return iterators that point to the first node/attribute, respectively; end and attributes_end return past-the-end iterator for node/attribute list, respectively - this iterator can’t be dereferenced, but decrementing it results in an iterator pointing to the last element in the list (except for empty lists, where decrementing past-the-end iterator results in undefined behavior). Past-the-end iterator is commonly used as a termination value for iteration loops (see sample below). If you want to get an iterator that points to an existing handle, you can construct the iterator with the handle as a single constructor argument, like so: xml_node_iterator(node). For xml_attribute_iterator, you’ll have to provide both an attribute and its parent node.

begin and end return equal iterators if called on null node; such iterators can’t be dereferenced. attributes_begin and attributes_end behave the same way. For correct iterator usage this means that child node/attribute collections of null nodes appear to be empty.

Both types of iterators have bidirectional iterator semantics (i.e. they can be incremented and decremented, but efficient random access is not supported) and support all usual iterator operations - comparison, dereference, etc. The iterators are invalidated if the node/attribute objects they’re pointing to are removed from the tree; adding nodes/attributes does not invalidate any iterators.

Here is an example of using iterators for document traversal (samples/traverse_iter.cpp):

The methods described above allow traversal of immediate children of some node; if you want to do a deep tree traversal, you’ll have to do it via a recursive function or some equivalent method. However, pugixml provides a helper for depth-first traversal of a subtree. In order to use it, you have to implement xml_tree_walker interface and to call traverse function:

The traversal is launched by calling traverse function on traversal root and proceeds as follows:

If begin, end or any of the for_each calls return false, the traversal is terminated and false is returned as the traversal result; otherwise, the traversal results in true. Note that you don’t have to override begin or end functions; their default implementations return true.

You can get the node’s depth relative to the traversal root at any point by calling depth function. It returns -1 if called from begin/end, and returns 0-based depth if called from for_each - depth is 0 for all children of the traversal root, 1 for all grandchildren and so on.

This is an example of traversing tree hierarchy with xml_tree_walker (samples/traverse_walker.cpp):

While there are existing functions for getting a node/attribute with known contents, they are often not sufficient for simple queries. As an alternative for manual iteration through nodes/attributes until the needed one is found, you can make a predicate and call one of find_ functions:

The predicate should be either a plain function or a function object which accepts one argument of type xml_attribute (for find_attribute) or xml_node (for find_child and find_node), and returns bool. The predicate is never called with null handle as an argument.

find_attribute function iterates through all attributes of the specified node, and returns the first attribute for which the predicate returned true. If the predicate returned false for all attributes or if there were no attributes (including the case where the node is null), null attribute is returned.

find_child function iterates through all child nodes of the specified node, and returns the first node for which the predicate returned true. If the predicate returned false for all nodes or if there were no child nodes (including the case where the node is null), null node is returned.

find_node function performs a depth-first traversal through the subtree of the specified node (excluding the node itself), and returns the first node for which the predicate returned true. If the predicate returned false for all nodes or if subtree was empty, null node is returned.

This is an example of using predicate-based functions (samples/traverse_predicate.cpp):

It is common to store data as text contents of some node - i.e. <node><description>This is a node</description></node>. In this case, <description> node does not have a value, but instead has a child of type node_pcdata with value "This is a node". pugixml provides a special class, xml_text, to work with such data. Working with text objects to modify data is described in the documentation for modifying document data; this section describes the access interface of xml_text.

You can get the text object from a node by using text() method:

If the node has a type node_pcdata or node_cdata, then the node itself is used to return data; otherwise, a first child node of type node_pcdata or node_cdata is used.

You can check if the text object is bound to a valid PCDATA/CDATA node by using it as a boolean value, i.e. if (text) { …​ } or if (!text) { …​ }. Alternatively you can check it by using the empty() method:

Given a text object, you can get the contents (i.e. the value of PCDATA/CDATA node) by using the following function:

In case text object is empty, the function returns an empty string - it never returns a null pointer.

If you need a non-empty string if the text object is empty, or if the text contents is actually a number or a boolean that is stored as a string, you can use the following accessors:

All of the above functions have the same semantics as similar xml_attribute members: they return the default argument if the text object is empty, they convert the text contents to a target type using the same rules and restrictions. You can refer to documentation for the attribute functions for details.

xml_text is essentially a helper class that operates on xml_node values. It is bound to a node of type node_pcdata or node_cdata. You can use the following function to retrieve this node:

Essentially, assuming text is an xml_text object, calling text.get() is equivalent to calling text.data().value().

This is an example of using xml_text object (samples/text.cpp):

If you need to get the document root of some node, you can use the following function:

This function returns the node with type node_document, which is the root node of the document the node belongs to (unless the node is null, in which case null node is returned).

While pugixml supports complex XPath expressions, sometimes a simple path handling facility is needed. There are two functions, for getting node path and for converting path to a node:

Node paths consist of node names, separated with a delimiter (which is / by default); also paths can contain self (.) and parent (..) pseudo-names, so that this is a valid path: "../../foo/./bar". path returns the path to the node from the document root, first_element_by_path looks for a node represented by a given path; a path can be an absolute one (absolute paths start with the delimiter), in which case the rest of the path is treated as document root relative, and relative to the given node. For example, in the following document: <a><b><c/></b></a>, node <c/> has path "a/b/c"; calling first_element_by_path for document with path "a/b" results in node <b/>; calling first_element_by_path for node <a/> with path "../a/./b/../." results in node <a/>; calling first_element_by_path with path "/a" results in node <a/> for any node.

In case path component is ambiguous (if there are two nodes with given name), the first one is selected; paths are not guaranteed to uniquely identify nodes in a document. If any component of a path is not found, the result of first_element_by_path is null node; also first_element_by_path returns null node for null nodes, in which case the path does not matter. path returns an empty string for null nodes.

pugixml does not record row/column information for nodes upon parsing for efficiency reasons. However, if the node has not changed in a significant way since parsing (the name/value are not changed, and the node itself is the original one, i.e. it was not deleted from the tree and re-added later), it is possible to get the offset from the beginning of XML buffer:

If the offset is not available (this happens if the node is null, was not originally parsed from a stream, or has changed in a significant way), the function returns -1. Otherwise it returns the offset to node’s data from the beginning of XML buffer in pugi::char_t units. For more information on parsing offsets, see parsing error handling documentation.

As discussed before, nodes can have name and value, both of which are strings. Depending on node type, name or value may be absent. node_document nodes do not have a name or value, node_element and node_declaration nodes always have a name but never have a value, node_pcdata, node_cdata, node_comment and node_doctype nodes never have a name but always have a value (it may be empty though), node_pi nodes always have a name and a value (again, value may be empty). In order to set node’s name or value, you can use the following functions:

Both functions try to set the name/value to the specified string, and return the operation result. The operation fails if the node can not have name or value (for instance, when trying to call set_name on a node_pcdata node), if the node handle is null, or if there is insufficient memory to handle the request. The provided string is copied into document managed memory and can be destroyed after the function returns (for example, you can safely pass stack-allocated buffers to these functions). The name/value content is not verified, so take care to use only valid XML names, or the document may become malformed.

This is an example of setting node name and value (samples/modify_base.cpp):

All attributes have name and value, both of which are strings (value may be empty). You can set them with the following functions:

Both functions try to set the name/value to the specified string, and return the operation result. The operation fails if the attribute handle is null, or if there is insufficient memory to handle the request. The provided string is copied into document managed memory and can be destroyed after the function returns (for example, you can safely pass stack-allocated buffers to these functions). The name/value content is not verified, so take care to use only valid XML names, or the document may become malformed.

In addition to string functions, several functions are provided for handling attributes with numbers and booleans as values:

The above functions convert the argument to string and then call the base set_value function. Integers are converted to a decimal form, floating-point numbers are converted to either decimal or scientific form, depending on the number magnitude, boolean values are converted to either "true" or "false".

For convenience, all set_value functions have the corresponding assignment operators:

These operators simply call the right set_value function and return the attribute they’re called on; the return value of set_value is ignored, so errors are ignored.

This is an example of setting attribute name and value (samples/modify_base.cpp):

Nodes and attributes do not exist without a document tree, so you can’t create them without adding them to some document. A node or attribute can be created at the end of node/attribute list or before/after some other node:

append_attribute and append_child create a new node/attribute at the end of the corresponding list of the node the method is called on; prepend_attribute and prepend_child create a new node/attribute at the beginning of the list; insert_attribute_after, insert_attribute_before, insert_child_after and insert_attribute_before add the node/attribute before or after the specified node/attribute.

Attribute functions create an attribute with the specified name; you can specify the empty name and change the name later if you want to. Node functions with the type argument create the node with the specified type; since node type can’t be changed, you have to know the desired type beforehand. Also note that not all types can be added as children; see below for clarification. Node functions with the name argument create the element node (node_element) with the specified name.

All functions return the handle to the created object on success, and null handle on failure. There are several reasons for failure:

Even if the operation fails, the document remains in consistent state, but the requested node/attribute is not added.

This is an example of adding new attributes/nodes to the document (samples/modify_add.cpp):

If you do not want your document to contain some node or attribute, you can remove it with one of the following functions:

remove_attribute removes the attribute from the attribute list of the node, and returns the operation result. remove_child removes the child node with the entire subtree (including all descendant nodes and attributes) from the document, and returns the operation result. Removing fails if one of the following is true:

Removing the attribute or node invalidates all handles to the same underlying object, and also invalidates all iterators pointing to the same object. Removing node also invalidates all past-the-end iterators to its attribute or child node list. Be careful to ensure that all such handles and iterators either do not exist or are not used after the attribute/node is removed.

If you want to remove the attribute or child node by its name, two additional helper functions are available:

These functions look for the first attribute or child with the specified name, and then remove it, returning the result. If there is no attribute or child with such name, the function returns false; if there are two nodes with the given name, only the first node is deleted. If you want to delete all nodes with the specified name, you can use code like this: while (node.remove_child("tool")) ;.

This is an example of removing attributes/nodes from the document (samples/modify_remove.cpp):

pugixml provides a special class, xml_text, to work with text contents stored as a value of some node, i.e. <node><description>This is a node</description></node>. Working with text objects to retrieve data is described in the documentation for accessing document data; this section describes the modification interface of xml_text.

Once you have an xml_text object, you can set the text contents using the following function:

This function tries to set the contents to the specified string, and returns the operation result. The operation fails if the text object was retrieved from a node that can not have a value and is not an element node (i.e. it is a node_declaration node), if the text object is empty, or if there is insufficient memory to handle the request. The provided string is copied into document managed memory and can be destroyed after the function returns (for example, you can safely pass stack-allocated buffers to this function). Note that if the text object was retrieved from an element node, this function creates the PCDATA child node if necessary (i.e. if the element node does not have a PCDATA/CDATA child already).

In addition to a string function, several functions are provided for handling text with numbers and booleans as contents:

The above functions convert the argument to string and then call the base set function. These functions have the same semantics as similar xml_attribute functions. You can refer to documentation for the attribute functions for details.

For convenience, all set functions have the corresponding assignment operators:

These operators simply call the right set function and return the attribute they’re called on; the return value of set is ignored, so errors are ignored.

This is an example of using xml_text object to modify text contents (samples/text.cpp):

With the help of previously described functions, it is possible to create trees with any contents and structure, including cloning the existing data. However since this is an often needed operation, pugixml provides built-in node/attribute cloning facilities. Since nodes and attributes do not exist without a document tree, you can’t create a standalone copy - you have to immediately insert it somewhere in the tree. For this, you can use one of the following functions:

These functions mirror the structure of append_child, prepend_child, insert_child_before and related functions - they take the handle to the prototype object, which is to be cloned, insert a new attribute/node at the appropriate place, and then copy the attribute data or the whole node subtree to the new object. The functions return the handle to the resulting duplicate object, or null handle on failure.

The attribute is copied along with the name and value; the node is copied along with its type, name and value; additionally attribute list and all children are recursively cloned, resulting in the deep subtree clone. The prototype object can be a part of the same document, or a part of any other document.

The failure conditions resemble those of append_child, insert_child_before and related functions, consult their documentation for more information. There are additional caveats specific to cloning functions:

This is an example with one possible implementation of include tags in XML (samples/include.cpp). It illustrates node cloning and usage of other document modification functions:

Sometimes instead of cloning a node you need to move an existing node to a different position in a tree. This can be accomplished by copying the node and removing the original; however, this is expensive since it results in a lot of extra operations. For moving nodes within the same document tree, you can use of the following functions instead:

These functions mirror the structure of append_copy, prepend_copy, insert_copy_before and insert_copy_after - they take the handle to the moved object and move it to the appropriate place with all attributes and/or child nodes. The functions return the handle to the resulting object (which is the same as the moved object), or null handle on failure.

The failure conditions resemble those of append_child, insert_child_before and related functions, consult their documentation for more information. There are additional caveats specific to moving functions:

pugixml provides several ways to assemble an XML document from other XML documents. Assuming there is a set of document fragments, represented as in-memory buffers, the implementation choices are as follows:

The first method is more convenient, but slower than the other two. The relative performance of append_copy and append_buffer depends on the buffer format - usually append_buffer is faster if the buffer is in native encoding (UTF-8 or wchar_t, depending on PUGIXML_WCHAR_MODE). At the same time it might be less efficient in terms of memory usage - the implementation makes a copy of the provided buffer, and the copy has the same lifetime as the document - the memory used by that copy will be reclaimed after the document is destroyed, but no sooner. Even deleting all nodes in the document, including the appended ones, won’t reclaim the memory.

append_buffer behaves in the same way as xml_document::load_buffer - the input buffer is a byte buffer, with size in bytes; the buffer is not modified and can be freed after the function returns.

Since append_buffer needs to append child nodes to the current node, it only works if the current node is either document or element node. Calling append_buffer on a node with any other type results in an error with status_append_invalid_root status.

If you want to save the whole document to a file, you can use one of the following functions:

These functions accept file path as its first argument, and also three optional arguments, which specify indentation and other output options (see Output options) and output data encoding (see Encodings). The path has the target operating system format, so it can be a relative or absolute one, it should have the delimiters of the target system, it should have the exact case if the target file system is case-sensitive, etc.

File path is passed to the system file opening function as is in case of the first function (which accepts const char* path); the second function either uses a special file opening function if it is provided by the runtime library or converts the path to UTF-8 and uses the system file opening function.

save_file opens the target file for writing, outputs the requested header (by default a document declaration is output, unless the document already has one), and then saves the document contents. If the file could not be opened, the function returns false. Calling save_file is equivalent to creating an xml_writer_file object with FILE* handle as the only constructor argument and then calling save; see Saving document via writer interface for writer interface details.

This is a simple example of saving XML document to file (samples/save_file.cpp):

To enhance interoperability pugixml provides functions for saving document to any object which implements C++ std::ostream interface. This allows you to save documents to any standard C++ stream (i.e. file stream) or any third-party compliant implementation (i.e. Boost Iostreams). Most notably, this allows for easy debug output, since you can use std::cout stream as saving target. There are two functions, one works with narrow character streams, another handles wide character ones:

save with std::ostream argument saves the document to the stream in the same way as save_file (i.e. with requested header and with encoding conversions). On the other hand, save with std::wstream argument saves the document to the wide stream with encoding_wchar encoding. Because of this, using save with wide character streams requires careful (usually platform-specific) stream setup (i.e. using the imbue function). Generally use of wide streams is discouraged, however it provides you with the ability to save documents to non-Unicode encodings, i.e. you can save Shift-JIS encoded data if you set the correct locale.

Calling save with stream target is equivalent to creating an xml_writer_stream object with stream as the only constructor argument and then calling save; see Saving document via writer interface for writer interface details.

This is a simple example of saving XML document to standard output (samples/save_stream.cpp):

All of the above saving functions are implemented in terms of writer interface. This is a simple interface with a single function, which is called several times during output process with chunks of document data as input:

In order to output the document via some custom transport, for example sockets, you should create an object which implements xml_writer interface and pass it to save function. xml_writer::write function is called with a buffer as an input, where data points to buffer start, and size is equal to the buffer size in bytes. write implementation must write the buffer to the transport; it can not save the passed buffer pointer, as the buffer contents will change after write returns. The buffer contains the chunk of document data in the desired encoding.

write function is called with relatively large blocks (size is usually several kilobytes, except for the last block that may be small), so there is often no need for additional buffering in the implementation.

This is a simple example of custom writer for saving document data to STL string (samples/save_custom_writer.cpp); read the sample code for more complex examples:

While the previously described functions save the whole document to the destination, it is easy to save a single subtree. The following functions are provided:

These functions have the same arguments with the same meaning as the corresponding xml_document::save functions, and allow you to save the subtree to either a C++ IOstream or to any object that implements xml_writer interface.

Saving a subtree differs from saving the whole document: the process behaves as if format_write_bom is off, and format_no_declaration is on, even if actual values of the flags are different. This means that BOM is not written to the destination, and document declaration is only written if it is the node itself or is one of node’s children. Note that this also holds if you’re saving a document; this example (samples/save_subtree.cpp) illustrates the difference:

All saving functions accept the optional parameter flags. This is a bitmask that customizes the output format; you can select the way the document nodes are printed and select the needed additional information that is output before the document contents.

These flags control the resulting tree contents:

These flags control the additional output information:

Additionally, there is one predefined option mask:

This is an example that shows the outputs of different output options (samples/save_options.cpp):

pugixml supports all popular Unicode encodings (UTF-8, UTF-16 (big and little endian), UTF-32 (big and little endian); UCS-2 is naturally supported since it’s a strict subset of UTF-16) and handles all encoding conversions during output. The output encoding is set via the encoding parameter of saving functions, which is of type xml_encoding. The possible values for the encoding are documented in Encodings; the only flag that has a different meaning is encoding_auto.

While all other flags set the exact encoding, encoding_auto is meant for automatic encoding detection. The automatic detection does not make sense for output encoding, since there is usually nothing to infer the actual encoding from, so here encoding_auto means UTF-8 encoding, which is the most popular encoding for XML data storage. This is also the default value of output encoding; specify another value if you do not want UTF-8 encoded output.

Also note that wide stream saving functions do not have encoding argument and always assume encoding_wchar encoding.

When you are saving the document using xml_document::save() or xml_document::save_file(), a default XML document declaration is output, if format_no_declaration is not specified and if the document does not have a declaration node. However, the default declaration is not customizable. If you want to customize the declaration output, you need to create the declaration node yourself.

Declaration node is a node with type node_declaration; it behaves like an element node in that it has attributes with values (but it does not have child nodes). Therefore setting custom version, encoding or standalone declaration involves adding attributes and setting attribute values.

This is an example that shows how to create a custom declaration node (samples/save_declaration.cpp):

Each XPath expression can have one of the following types: boolean, number, string or node set. Boolean type corresponds to bool type, number type corresponds to double type, string type corresponds to either std::string or std::wstring, depending on whether wide character interface is enabled, and node set corresponds to xpath_node_set type. There is an enumeration, xpath_value_type, which can take the values xpath_type_boolean, xpath_type_number, xpath_type_string or xpath_type_node_set, accordingly.

Because an XPath node can be either a node or an attribute, there is a special type, xpath_node, which is a discriminated union of these types. A value of this type contains two node handles, one of xml_node type, and another one of xml_attribute type; at most one of them can be non-null. The accessors to get these handles are available:

XPath nodes can be null, in which case both accessors return null handles.

Note that as per XPath specification, each XPath node has a parent, which can be retrieved via this function:

parent function returns the node’s parent if the XPath node corresponds to xml_node handle (equivalent to node().parent()), or the node to which the attribute belongs to, if the XPath node corresponds to xml_attribute handle. For null nodes, parent returns null handle.

Like node and attribute handles, XPath node handles can be implicitly cast to boolean-like object to check if it is a null node, and also can be compared for equality with each other.

You can also create XPath nodes with one of the three constructors: the default constructor, the constructor that takes node argument, and the constructor that takes attribute and node arguments (in which case the attribute must belong to the attribute list of the node). The constructor from xml_node is implicit, so you can usually pass xml_node to functions that expect xpath_node. Apart from that you usually don’t need to create your own XPath node objects, since they are returned to you via selection functions.

XPath expressions operate not on single nodes, but instead on node sets. A node set is a collection of nodes, which can be optionally ordered in either a forward document order or a reverse one. Document order is defined in XPath specification; an XPath node is before another node in document order if it appears before it in XML representation of the corresponding document.

Node sets are represented by xpath_node_set object, which has an interface that resembles one of sequential random-access containers. It has an iterator type along with usual begin/past-the-end iterator accessors:

And it also can be iterated via indices, just like std::vector:

All of the above operations have the same semantics as that of std::vector: the iterators are random-access, all of the above operations are constant time, and accessing the element at index that is greater or equal than the set size results in undefined behavior. You can use both iterator-based and index-based access for iteration, however the iterator-based one can be faster.

The order of iteration depends on the order of nodes inside the set; the order can be queried via the following function:

type function returns the current order of nodes; type_sorted means that the nodes are in forward document order, type_sorted_reverse means that the nodes are in reverse document order, and type_unsorted means that neither order is guaranteed (nodes can accidentally be in a sorted order even if type() returns type_unsorted). If you require a specific order of iteration, you can change it via sort function:

Calling sort sorts the nodes in either forward or reverse document order, depending on the argument; after this call type() will return type_sorted or type_sorted_reverse.

Often the actual iteration is not needed; instead, only the first element in document order is required. For this, a special accessor is provided:

This function returns the first node in forward document order from the set, or null node if the set is empty. Note that while the result of the node does not depend on the order of nodes in the set (i.e. on the result of type()), the complexity does - if the set is sorted, the complexity is constant, otherwise it is linear in the number of elements or worse.

While in the majority of cases the node set is returned by XPath functions, sometimes there is a need to manually construct a node set. For such cases, a constructor is provided which takes an iterator range (const_iterator is a typedef for const xpath_node*), and an optional type:

The constructor copies the specified range and sets the specified type. The objects in the range are not checked in any way; you’ll have to ensure that the range contains no duplicates, and that the objects are sorted according to the type parameter. Otherwise XPath operations with this set may produce unexpected results.

If you want to select nodes that match some XPath expression, you can do it with the following functions:

select_nodes function compiles the expression and then executes it with the node as a context node, and returns the resulting node set. select_node returns only the first node in document order from the result, and is equivalent to calling select_nodes(query).first(). If the XPath expression does not match anything, or the node handle is null, select_nodes returns an empty set, and select_node returns null XPath node.

If exception handling is not disabled, both functions throw xpath_exception if the query can not be compiled or if it returns a value with type other than node set; see Error handling for details.

While compiling expressions is fast, the compilation time can introduce a significant overhead if the same expression is used many times on small subtrees. If you’re doing many similar queries, consider compiling them into query objects (see Using query objects for further reference). Once you get a compiled query object, you can pass it to select functions instead of an expression string:

If exception handling is not disabled, both functions throw xpath_exception if the query returns a value with type other than node set.

This is an example of selecting nodes using XPath expressions (samples/xpath_select.cpp):

When you call select_nodes with an expression string as an argument, a query object is created behind the scenes. A query object represents a compiled XPath expression. Query objects can be needed in the following circumstances:

Query objects correspond to xpath_query type. They are immutable and non-copyable: they are bound to the expression at creation time and can not be cloned. If you want to put query objects in a container, allocate them on heap via new operator and store pointers to xpath_query in the container.

You can create a query object with the constructor that takes XPath expression as an argument:

The expression is compiled and the compiled representation is stored in the new query object. If compilation fails, xpath_exception is thrown if exception handling is not disabled (see Error handling for details). After the query is created, you can query the type of the evaluation result using the following function:

You can evaluate the query using one of the following functions:

All functions take the context node as an argument, compute the expression and return the result, converted to the requested type. According to XPath specification, value of any type can be converted to boolean, number or string value, but no type other than node set can be converted to node set. Because of this, evaluate_boolean, evaluate_number and evaluate_string always return a result, but evaluate_node_set and evaluate_node result in an error if the return type is not node set (see Error handling).

Note that evaluate_string function returns the STL string; as such, it’s not available in PUGIXML_NO_STL mode and also usually allocates memory. There is another string evaluation function:

This function evaluates the string, and then writes the result to buffer (but at most capacity characters); then it returns the full size of the result in characters, including the terminating zero. If capacity is not 0, the resulting buffer is always zero-terminated. You can use this function as follows:

This is an example of using query objects (samples/xpath_query.cpp):