XML Parsing and Serialization in C++
With libstudxml

Copyright © 2013-2014 Code Synthesis Tools CC. Permission is granted to copy, distribute and/or modify this document under the terms of the MIT license.

Revision 1.0, May 2014

This revision of the document describes libstudxml 1.0.0.


Table of Contents

About This Document
1Terminology
2Low-Level API
3High-Level API
4Object Persistence
5Inheritance
6Implementation Notes

About This Document

This document is based on the presentation given by Boris Kolpackov at the C++Now 2014 conference where libstudxml was first made publicly available. Its goal is to introduce a new, modern C++ API for XML by showing how to handle the most common use cases. Compared to the talk, this introduction omits some of the discussion relevant to XML in general and its handling in C++. It also provides more complete code examples that would not fit onto slides during the presentation. If, however, you would like to get a more complete picture of the "state of XML in C++", then you may prefer to first watch the video of the talk.

While this document uses some C++11 features in the examples, the library itself can be used in C++98 applications as well.

Terminology

Before we begin, let's define a few terms to make sure we are on the same page.

When we say "XML format" that is a bit loose. XML is actually a meta-format that we specialize for our needs. That is, we decide what element and attribute names we will use, which elements will be valid where, what they will mean, and so on. This specialization of XML to a specific format is called an XML Vocabulary.

Often, but not always, when we parse XML, we store extracted data in the application's memory. Usually, we would create classes specific to our XML vocabulary. For example, if we have an element called person then we may create a C++ class also called person. we will call such classes an Object Model.

The content of an element in XML can be empty, text, nested elements, or a mixture of the two:

<empty name="a" id="1"/>

<simple name="b" id="2">text<simple/>

<complex name="c" id="3">
  <nested>...</nested>
  <nested>...</nested>
<complex/>

<mixed name="d" id="4">
  te<nested>...</nested>
  x
  <nested>...</nested>t
<mixed/>
  

These are called the empty, simple, complex, and mixed content models, respectively.

Low-Level API

libstudxml provides the streaming XML pull parser and streaming XML serializer. The parser is a conforming, non-validating XML 1.0 implementation (see Implementation Notes for details). The application character encoding (that is, the encoding used in the application's memory) for both parser and serializer is UTF-8. The output encoding of the serializer is UTF-8 as well. The parser supports UTF-8, UTF-16, ISO-8859-1, and US-ASCII input encodings.

#include <xml/parser>

namespace xml
{
  class parser;
}
  
#include <xml/serializer>

namespace xml
{
  class serializer;
}
  

C++ is often used to implement XML converters and filters, especially where speed is a concern. Such applications require the lowest-level API with minimum overhead. So we will start there (see the roundtrip example in the libstudxml distribution).

class parser
{
  typedef unsigned short feature_type;

  static const feature_type receive_elements;
  static const feature_type receive_characters;
  static const feature_type receive_attributes;
  static const feature_type receive_namespace_decls;

  static const feature_type receive_default =
    receive_elements |
    receive_characters |
    receive_attributes;

  parser (std::istream&,
          const std::string& input_name,
          feature_type = receive_default);
  ...
};
  

The parser constructor takes three arguments: the stream to parse, input name that is used in diagnostics to identify the document being parsed, and the list of events we want the parser to report.

As an example of an XML filter, let's write one that removes a specific attribute from the document, say id. The first step in our filter would then be to create the parser instance:

int main (int argc, char* argv[])
{
  ...

  try
  {
    using namespace xml;

    ifstream ifs (argv[1]);
    parser p (ifs, argv[1]);

    ...
  }
  catch (const xml::parsing& e)
  {
    cerr << e.what () << endl;
    return 1;
  }
}
  

Here we also see how to handle parsing errors. So far so good. Let's see the next piece of the API.

class parser
{
  enum event_type
  {
    start_element,
    end_element,
    start_attribute,
    end_attribute,
    characters,
    start_namespace_decl,
    end_namespace_decl,
    eof
  };

  event_type next ();
};
  

We call the next() function when we are ready to handle the next piece of XML. And now we can implement our filter a bit further:

parser p (ifs, argv[1]);

for (parser::event_type e (p.next ());
     e != parser::eof;
     e = p.next ())
{
  switch (e)
  {
  case parser::start_element:
    ...
  case parser::end_element:
    ...
  case parser::start_attribute:
    ...
  case parser::end_attribute:
    ...
  case parser::characters:
    ...
  }
}
  

In C++11 we can use the range-based for loop to tidy things up a bit:

parser p (ifs, argv[1]);

for (parser::event_type e: p)
{
  switch (e)
  {
    ...
  }
}
  

The next piece of the API puzzle:

class parser
{
  const std::string& name () const;
  const std::string& value () const;

  unsigned long long line () const;
  unsigned long long column () const;
};
  

The name() accessor returns the name of the current element or attribute. The value() function returns the text of the characters event for an element or attribute. The line() and column() accessors return the current position in the document. Here is how we could print all the element positions for debugging:

switch (e)
{
case parser::start_element:
  cerr << p.line () << ':' << p.column () << ": start "
       << p.name () << endl;
  break;
case parser::end_element:
  cerr << p.line () << ':' << p.column () << ": end "
       << p.name () << endl;
  break;
}
  

We have now seen enough of the parsing side to complete our filter. What's missing is the serialization. So let's switch to that for a moment:

class serializer
{
  serializer (std::ostream&,
              const std::string& output_name,
              unsigned short indentation = 2);

  ...
};
  

The constructor is pretty similar to the parser's. The indentation argument specifies the number of indentation spaces that should be used for pretty-printing. We can disable it by passing 0.

Now we can create the serializer instance for our filter:

int main (int argc, char* argv[])
{
  ...

  try
  {
    using namespace xml;

    ifstream ifs (argv[1]);
    parser p (ifs, argv[1]);
    serializer s (cout, "output", 0);

    ...
  }
  catch (const xml::parsing& e)
  {
    cerr << e.what () << endl;
    return 1;
  }
  catch (const xml::serialization& e)
  {
    cerr << e.what () << endl;
    return 1;
  }
}
  

Notice that we have also added an exception handler for the serialization exception. Instead of handling the parsing and serialization exceptions separately, we can catch just xml::exception, which is a common base for the other two:

int main (int argc, char* argv[])
{
  try
  {
    ...
  }
  catch (const xml::exception& e)
  {
    cerr << e.what () << endl;
    return 1;
  }
}
  

The next chunk of the serializer API:

class serializer
{
  void start_element (const std::string& name);
  void end_element ();

  void start_attribute (const std::string& name);
  void end_attribute ();

  void characters (const std::string& value);
};
  

Everything should be pretty self-explanatory here. And we have now seen enough to finish our filter:

parser p (ifs, argv[1]);
serializer s (cout, "output", 0);

bool skip (false);

for (parser::event_type e: p)
{
  switch (e)
  {
  case parser::start_element:
    {
      s.start_element (p.name ());
      break;
    }
  case parser::end_element:
    {
      s.end_element ();
      break;
    }
  case parser::start_attribute:
    {
      if (p.name () == "id")
        skip = true;
      else
        s.start_attribute (p.name ());
      break;
    }
  case parser::end_attribute:
    {
      if (skip)
        skip = false;
      else
        s.end_attribute ();
      break;
    }
  case parser::characters:
    {
      if (!skip)
        s.characters (p.value ());
      break;
    }
  }
}
  

Do you see any problems with our filter? Well, one problem is that this implementation doesn't handle XML namespaces. Let's see how we can fix this. The first issue is with the element and attribute names. When namespaces are used, those may be qualified. libstudxml uses the qname class to represent such names:

#include <xml/qname>

namespace xml
{
  class qname
  {
  public:
    qname ();
    qname (const std::string& name);
    qname (const std::string& namespace_,
           const std::string& name);

    const std::string& namespace_ () const;
    const std::string& name () const;
  };
}
  

The parser, in addition to the name() accessor also has qname() which returns the potentially qualified name. Similarly, the start_element() and start_attribute() functions in the serializer are overloaded to accept qname:

class parser
{
  const qname& qname () const;
};

class serializer
{
  void start_element (const qname&);
  void start_attribute (const qname&);
};
  

The first thing we need to do to make our filter namespace-aware is to use qualified names instead of the local ones. This one is easy:

switch (e)
{
case parser::start_element:
  {
    s.start_element (p.qname ());
    break;
  }
case parser::start_attribute:
  {
    if (p.qname () == "id") // Unqualified name.
      skip = true;
    else
      s.start_attribute (p.qname ());
    break;
  }
}
  

There is, however, another thing that we have to do. Right now our code does not propagate the namespace-prefix mappings from the input document to the output. At the moment, where the input XML might have meaningful prefixes assigned to namespaces, the output will have automatically generated ones like g1, g2, and so on.

To fix this, first we need to tell the parser to report to us namespace-prefix mappings, called namespace declarations in XML:

parser p (ifs,
          argv[1]
          parser::receive_default |
          parser::receive_namespace_decls);
  

We then also need to propagate this information to the serializer by handling the start_namespace_decl event:

for (...)
{
  switch (e)
  {
    ...

  case parser::start_namespace_decl:
    s.namespace_decl (p.namespace_ (), p.prefix ());
    break;

    ...
  }
}
  

Well, that wasn't too bad.

High-Level API

So that was pretty low level XML work where we didn't care about the semantics of the stored data, or, in fact the XML vocabulary that we dealt with.

However, this API will quickly become tedious once we try to handle a specific XML vocabulary and do something useful with the stored data. Why is that? There are several areas where we could use some help:

Let's examine each area using our object position vocabulary as a test case (see the processing example in the libstudxml distribution).

<object id="123">
  <name>Lion's Head</name>
  <type>mountain</type>

  <position lat="-33.8569" lon="18.5083"/>
  <position lat="-33.8568" lon="18.5083"/>
  <position lat="-33.8568" lon="18.5082"/>
</object>
  

If you cannot assume the XML you are parsing is valid, and you generally shouldn't, then you will quickly realize that the biggest pain in dealing with XML is making sure that what we got is actually valid.

This stuff is pervasive. What if the root element is spelled wrong? Maybe the id attribute is missing? Or there is some stray text before the name element? Things can be broken in an infinite number of ways.

To illustrate this point, here is the parsing code of just the root element with proper error handling:

parser p (ifs, argv[1]);

if (p.next () != parser::start_element ||
    p.qname () != "object")
{
  // error
}

...

if (p.next () != parser::end_element) // object
{
  // error
}
  

Not very pretty. To help with this, the parser API provides the next_expect() function:

class parser
{
  void next_expect (event_type);
  void next_expect (event_type, const std::string& name);
};
  

This function gets the next event and makes sure it is what's expected. If not, it throws an appropriate parsing exception. This simplifies our root element parsing quite a bit:

parser p (ifs, argv[1]);

p.next_expect (parser::start_element, "object");
...
p.next_expect (parser::end_element); // object
  

Let's now take the next step and try to handle the id attribute. According to what we have seen so far, it will look something along these lines:

p.next_expect (parser::start_element, "object");

p.next_expect (parser::start_attribute, "id");
p.next_expect (parser::characters);
cout << "id: " << p.value () << endl;
p.next_expect (parser::end_attribute);

...

p.next_expect (parser::end_element); // object
  

Not too bad but there is a bit of a problem. What if our object element had several attributes? The order of attributes in XML is arbitrary so we should be prepared to get them in any order. This fact complicates our attribute parsing code quite a bit:

while (p.next () == parser::start_attribute)
{
  if (p.qname () == "id")
  {
    p.next_expect (parser::characters);
    cout << "id: " << p.value () << endl;
  }
  else if (...)
  {
  }
  else
  {
    // error: unknown attribute
  }

  p.next_expect (parser::end_attribute);
}
  

There is also a bug in this version. Can you see it? We now don't make sure that the id attribute was actually specified.

If you think about it, at this level, it is actually not that convenient to receive attributes as events. In fact, a map of attributes would be much more usable.

Remember we talked about the parser features that specify which events we want to see:

class parser
{
  static const feature_type receive_elements;
  static const feature_type receive_characters;
  static const feature_type receive_attributes;

  ...
};
  

Well, in reality, there is no receive_attributes. Rather, there are these two options:

class parser
{
  static const feature_type receive_attributes_map;
  static const feature_type receive_attributes_event;

  ...
};
  

That is, we can ask the parser to send us attributes as events or as a map. And the default is to send them as a map.

In case of a map, we have the following attribute access API to work with:

class parser
{
  const std::string& attribute (const std::string& name) const;

  std::string attribute (const std::string& name,
                         const std::string& default_value) const;

  bool attribute_present (const std::string& name) const;
};
  

If the attribute is not found, then the version without the default value throws an appropriate parsing exception while the version with the default value returns that value. There are also the qname versions of these functions.

Let's see how this simplifies our code:

p.next_expect (parser::start_element, "object");

cout << "id: " << p.attribute ("id") << endl;

...

p.next_expect (parser::end_element); // object
  

Much better.

If the id attribute is not present, then we get an exception. But what happens if we have a stray attribute in our document? The attribute map is magical in this sense. After the end_element event for the object element the parser will examine the attribute map. If there is an attribute that hasn't been retrieved with one of the attribute access functions, then the parser will throw the unexpected attribute exception.

Error handling out of the way, the next thing that will annoy us is data extractions. In XML everything is text. While our id value is an integer, XML stores it as text and the low-level API returns it to us as text. To help with this the parser provides the following data extraction functions:

class parser
{
  template <typename T>
  T value () const;

  template <typename T>
  T attribute (const std::string& name) const;

  template <typename T>
  T attribute (const std::string& name,
               const T& default_value) const;
};
  

Now we can get the id as an integer without much fuss:

p.next_expect (parser::start_element, "object");

unsigned int id = p.attribute<unsigned int> ("id");

...

p.next_expect (parser::end_element); // object
  

Ok, let's try to parse our vocabulary a bit further:

p.next_expect (parser::start_element, "object");
unsigned int id = p.attribute<unsigned int> ("id");

p.next_expect (parser::start_element, "name");

...

p.next_expect (parser::end_element); // name

p.next_expect (parser::end_element); // object
  

Here is the part of the document that we are parsing:

<object id="123">
  <name>Lion's Head</name>
  

What do you think, is everything alright with our code? When we try to parse our document, we will get an exception here:

p.next_expect (parser::start_element, "name");
  

Any idea why? Let's try to print the event that we get:

// p.next_expect (parser::start_element, "name");
cerr << p.next () << endl;
  

We expect start_element but get characters! Wait a minute, but there are characters after object and before name. There is a newline and two spaces that are replaced with hashes for illustration here:

<object id="123">#
##<name>Lion's Head</name>
  

If you go to a forum or a mailing list for any XML parser, this will be the most common question. Why do I get text when I should clearly get an element!?

The reason why we get this whitespace text is because the parser has no idea whether it is significant or not. The significance of whitespaces is determined by the XML content model that we talked about earlier. Here is the table:

#include <xml/content>

namespace xml
{
  enum class content
  {          //  element   characters  whitespaces
    empty,   //    no          no        ignored
    simple,  //    no          yes       preserved
    complex, //    yes         no        ignored
    mixed    //    yes         yes       preserved
  };
}
  

In empty content neither nested elements nor characters are allowed with whitespaces ignored. Simple content allows no nested elements with whitespaces preserved. Complex content allows nested elements only with whitespaces which are ignored. Finally, the mixed content allows anything in any order with everything preserved.

If we specify the content model for an element, then the parser will do automatic whitespace processing for us:

class parser
{
  void content (content);
};
  

That is, in empty and complex content, whitespaces will be silently ignored. By knowing the content model, the parser also has a chance to do more error handling for us. It will automatically throw appropriate exceptions if there are nested elements in empty or simple content or non-whitespace characters in complex content.

Ok, let's now see how we can take advantage of this feature in our code:

p.next_expect (parser::start_element, "object");
p.content (content::complex);

unsigned int id = p.attribute<unsigned int> ("id");

p.next_expect (parser::start_element, "name"); // Ok.

...

p.next_expect (parser::end_element); // name

p.next_expect (parser::end_element); // object
  

Now whitespaces are ignored and everything works as we expected. Here is how we can parse the content of the name element:

p.next_expect (parser::start_element, "name");
p.content (content::simple);

p.next_expect (parser::characters);
string name = p.value ();

p.next_expect (parser::end_element); // name
  

As you can see, parsing a simple content element is quite a bit more involved compared to getting a value of an attribute. Element markup also has a higher overhead in the resulting XML. That's why in our case it would have been wiser to make name and type attributes.

But if we are stuck with a lot of simple content elements, then the parser provides the following helper functions:

class parser
{
  std::string element ();

  template <typename T>
  T element ();

  std::string element (const std::string& name);

  template <typename T>
  T element (const std::string& name);

  std::string element (const std::string& name,
                       const std::string& default_value);

  template <typename T>
  T element (const std::string& name,
             const T& default_value);
};
  

The first two assume that you have already handled the start_element event. They should be used if the element also has attributes. The other four parse the complete element. Overloaded qname versions are also provided.

Here is how we can simplify our parsing code thanks to these functions:

p.next_expect (parser::start_element, "object");
p.content (content::complex);

unsigned int id = p.attribute<unsigned int> ("id");
string name = p.element ("name");

p.next_expect (parser::end_element); // object
  

For the type element we would like to use this enum class:

enum class object_type
{
  building,
  mountain,
  ...
};
  

The parsing code is similar to the name element. Now we use the data extracting version of the element() function:

object_type type = p.element<object_type> ("type");
  

Except that this won't compile. The parser doesn't know how to convert the text representation to our enum. By default the parser will try to use the iostream extraction operator but we haven't provided any.

We can provide conversion code specifically for XML by specializing the value_traits class template:

namespace xml
{
  template <>
  struct value_traits<object_type>
  {
    static object_type
    parse (std::string, const parser&)
    {
      ...
    }

    static std::string
    serialize (object_type, const serializer&)
    {
      ...
    }
  };
}
  

The last bit that we need to handle is the position elements. The interesting part here is how to stop without going too far since there can be several of them. To help with this task the parser allows us to peek into the next event:

p.next_expect (parser::start_element, "object");
p.content (content::complex);
...

do
{
  p.next_expect (parser::start_element, "position");
  p.content (content::empty);

  float lat = p.attribute<float> ("lat");
  float lon = p.attribute<float> ("lon");

  p.next_expect (parser::end_element);

} while (p.peek () == parser::start_element);

p.next_expect (parser::end_element); // object
  

Do you see anything else that we can improve? Actually, there is one thing. Look at the next_expect() calls in the above code. They are both immediately followed by the setting of the content model. We can tidy this up a bit by passing the content model as a third argument to next_expect(). This even reads like prose: "Next we expect the start of an element called position that shall have empty content."

Here is the complete, production-quality parsing code for our XML vocabulary. 13 lines. With validation and everything:

parser p (ifs, argv[1]);

p.next_expect (parser::start_element, "object", content::complex);

unsigned int id = p.attribute<unsigned int> ("id");
string name = p.element ("name");
object_type type = p.element<object_type> ("type");

do
{
  p.next_expect (parser::start_element, "position", content::empty);

  float lat = p.attribute<float> ("lat");
  float lon = p.attribute<float> ("lon");

  p.next_expect (parser::end_element); // position
} while (p.peek () == parser::start_element)

p.next_expect (parser::end_element); // object
  

So that was the high-level parsing API. Let's now catch up with the corresponding additions to the serializer.

Similar to parsing, calling start_attribute(), characters(), and then end_attribute() might not be convenient. Instead we can add an attribute with a single call:

class serializer
{
  void attribute (const std::string& name,
                  const std::string& value);

  void element (const std::string& value);

  void element (const std::string& name,
                const std::string& value);
};
  

The same works for elements with simple content. The first version finishes the element that we have started, while the second writes the complete element. There are also the qname versions of these functions that are not shown.

Instead of strings we can also serialize value types. This uses the same value_traits specialization mechanism that we have used for parsing:

class serializer
{
  template <typename T>
  void attribute (const std::string& name,
                  const T& value);

  template <typename T>
  void element (const T& value);

  template <typename T>
  void element (const std::string& name,
                const T& value);

  template <typename T>
  void characters (const T& value);
};
  

Let's now see now how we can serialize a complete sample document for our object position vocabulary using this high-level API:

serializer s (cout, "output");

s.start_element ("object");

s.attribute ("id", 123);
s.element ("name", "Lion's Head");
s.element ("type", object_type::mountain);

for (...)
{
  s.start_element ("position");

  float lat (...), lon (...);

  s.attribute ("lat", lat);
  s.attribute ("lon", lon);

  s.end_element (); // position
}

s.end_element (); // object
  

Pretty straightforward stuff.

Object Persistence

So far we have used our API to first implement a filter that doesn't really care about the data and then an application that processes the data without creating any kind of object model. Let's now try to handle the other end of the spectrum: objects that know how to persist themselves into XML (see the persistence example in the libstudxml distribution).

But before we continue, let's fix our XML to be slightly more idiomatic. That is we make name and type to be attributes rather than elements:

<object name="Lion's Head" type="mountain" id="123">
  <position lat="-33.8569" lon="18.5083"/>
  <position lat="-33.8568" lon="18.5083"/>
  <position lat="-33.8568" lon="18.5082"/>
</object>
  

Generally, the API works best with idiomatic XML and will nudge you gently in that direction with minor inconveniences.

For this vocabulary, the object model might look like this:

enum class object_type {...};

class position
{
  ...

  float lat_;
  float lon_;
};

class object
{
  ...

  std::string name_;
  object_type type_;
  unsigned int id_;
  std::vector<position> positions_;
};
  

Here I omit sensible constructors, accessors and modifiers that our classes would probably have.

Let me also mention that what I am going to show next is what I believe is the sensible structure for XML persistence using this API. But that doesn't mean it is the only way. For example, we are going to do parsing in a constructor:

class position
{
  position (xml::parser&);

  void
  serialize (xml::serializer&) const;

  ...
};

class object
{
  object (xml::parser&);

  void
  serialize (xml::serializer&) const;

  ...
};
  

But you may prefer to first create an instance, say with the default constructor, and then have a separate function do the parsing. There is nothing wrong with this approach.

Let's start with the position constructor. Here, we are immediately confronted with this choice: do we parse the start and end element events in position or expect our caller to handle them.

I suggest that we let our caller do this. We may have different elements in our vocabulary that use the same position type. If we assume the element name in the constructor, then we won't be able to use the same class for all these elements. We will see the second advantage of this arrangement in a moment, when we deal with inheritance. But, if you have a simple model with one-to-one mapping between types and elements and no inheritance, then there is nothing wrong with going the other route.

position::
position (parser& p)
  : lat_ (p.attribute<float> ("lat")),
    lon_ (p.attribute<float> ("lon"))
{
  p.content (content::empty);
}
  

Ok, nice and clean so far. Let's look at the object constructor:

object::
object (parser& p)
  : name_ (p.attribute ("name")),
    type_ (p.attribute<object_type> ("type")),
    id_ (p.attribute<unsigned int> ("id"))
{
  p.content (content::complex);

  do
  {
    p.next_expect (parser::start_element, "position");
    positions_.push_back (position (p));
    p.next_expect (parser::end_element);

  } while (p.peek () == parser::start_element);
}
  

The only mildly interesting line here is where we call the position constructor to parse the content of the nested elements.

Before we look into serialization, let me also mention one other thing. In our vocabulary all the attributes are required but it is quite common to have optional attributes. The API functions with default values make it really convenient to handle such attributes in the initializer lists.

Let's say the type attribute is optional. Then we could do this:

object::
object (parser& p)
  : ...
    type_ (p.attribute ("type", object_type::other))
    ...
  

We use the same arrangement for serialization, that is, the containing object starts and ends the element allowing us to reuse the same type for different elements:

void position::serialize (serializer& s) const
{
  s.attribute ("lat", lat_);
  s.attribute ("lon", lon_);
}

void object::serialize (serializer& s) const
{
  s.attribute ("name", name_);
  s.attribute ("type", type_);
  s.attribute ("id", id_);

  for (const auto& p: positions_)
  {
    s.start_element ("position");
    p.serialize (s);
    s.end_element ();
  }
}
  

Ok, also nice and tidy.

There is one thing, however, that is not so nice: the start of the parser or serializer. Here is the code:

parser p (ifs, argv[1]);
p.next_expect (parser::start_element, "object");
object o (p);
p.next_expect (parser::end_element);

serializer s (cout, "output");
s.start_element ("object");
o.serialize (s);
s.end_element ();
  

Remember, we made the caller responsible for handling the start and end of the element. This works beautifully inside the object model but not so much in the client code. What we would like to see instead is this:

parser p (ifs, argv[1]);
object o (p);

serializer s (cout, "output");
o.serialize (s);
  

The main reason for choosing this structure was the ability to reuse the same type for different elements. The other reason was inheritance which we haven't gotten to yet. If we think about it, it is very unlikely for a class corresponding to the root of our vocabulary to also be used inside as a local element. I can't remember ever seeing a vocabulary like this.

So what we can do here is make an exception: the root type of our object model handles the top-level element. Here is the parser:

object::
object (parser& p)
{
  p.next_expect (
    parser::start_element, "object", content::complex);

  name_ = p.attribute ("name");
  type_ = p.attribute<object_type> ("type");
  id_ = p.attribute<unsigned int> ("id");

  ...

  p.next_expect (parser::end_element);
}
  

And here is the serializer:

void object::
serialize (serializer& s) const
{
  s.start_element ("object");

  ...

  s.end_element ();
}
  

The only minor drawback of going this route is that we can no longer parse attributes in the initializer list for the root object.

Inheritance

So far we have had a smooth sailing with the streaming approach but things get a bit bumpy once we start dealing with inheritance. This is normally where the in-memory approach has its day.

Say we have elevated-object which adds the units attribute and the elevation elements. Here is the XML:

<elevated-object name="Lion's Head" type="mountain"
                 units="m" id="123">
  <position lat="-33.8569" lon="18.5083"/>
  <position lat="-33.8568" lon="18.5083"/>
  <position lat="-33.8568" lon="18.5082"/>

  <elevation val="668.9"/>
  <elevation val="669"/>
  <elevation val="669.1"/>
</elevated-object>
  

And here is the object model:

enum class units {...};

class elevation {...};

class elevated_object: public object
{
  ...

  units units_;
  std::vector<elevation> elevations_;
};
  

Streaming assumes linearity. We start an element, add some attributes, add some nested elements, and end the element. In contrast, with an in-memory approach we can add some attributes, then add some nested elements, then go back and add more attributes. This kind of back and forth is exactly what inheritance often requires. So this is a bit of problem for us.

Consider the elevated_object constructor:

elevated_object::
elevated_object (parser& p)
  : object (p),
    units_ (p.attribute<units> ("units"))
{
  do
  {
    p.next_expect (parser::start_element, "elevation");
    elevations_.push_back (elevation (p));
    p.next_expect (parser::end_element);

  } while (p.peek () == parser::start_element &&
           p.name () == "elevation")
}
  

Note that here I assume we went back to our original architecture where the caller handles the start and end of the element (this is the other advantage of this architecture: it allows us to reuse base parsing and serialization code in derived classes).

So we would like to reuse the parsing code from object so we call the base constructor first.

Then we parse the derived attribute and elements. Do you see the problem? The object constructor will parse its attributes and then move on to nested elements. When this constructor returns, we need to go back to parsing attributes! This is not something that a streaming approach would normally allow.

To resolve this, the lifetime of the attribute map was extended until after the end_element event. That is, we can access attributes any time we are at the element's level. As a result, the above code just works.

We have the same problem in serialization. Let's say we write the straightforward code like this:

void elevated_object::
serialize (serializer& s) const
{
  object::serialize (s);

  s.attribute ("units", units_);

  for (const auto& e: elevations_)
  {
    s.start_element ("elevation");
    e.serialize (s);
    s.end_element ();
  }
}
  

This is not going to work since we will try to add the units attribute after the nested position elements have already been written.

To handle inheritance in serialization we have to split the serialize() function into two. One serializes the attributes while the other — content:

void object::
serialize_attributes (serializer& s) const
{
  s.attribute ("name", name_);
  s.attribute ("type", type_);
  s.attribute ("id", id_);
}

void object::
serialize_content (serializer& s) const
{
  for (const auto& p: positions_)
  {
    s.start_element ("position");
    p.serialize (s);
    s.end_element ();
  }
}
  

The serialize() function then simply calls these two in the correct order.

void object::
serialize (serializer& s) const
{
  serialize_attributes (s);
  serialize_content (s);
}
  

I bet you can guess what the elevated_object's implementation looks like:

void elevated_object::
serialize_attributes (serializer& s) const
{
  object::serialize_attributes (s);
  s.attribute ("units", units_);
}

void elevated_object::
serialize_content (serializer& s) const
{
  object::serialize_content (s);

  for (const auto& e: elevations_)
  {
    s.start_element ("elevation");
    e.serialize (s);
    s.end_element ();
  }
}
  

The serialize() function for elevated_object is exactly the same:

void elevated_object::
serialize (serializer& s) const
{
  serialize_attributes (s);
  serialize_content (s);
}
  

Implementation Notes

libstudxmlis an open source (MIT license), portable (autotools and VC++ projects provided), and external dependency-free implementation.

It provides a conforming, non-validating XML 1.0 parser by using the mature and tested Expat XML parser. libstudxml includes the Expat source code (also distributed under the MIT license) as an implementation detail. However, you can link to an external Expat library if you prefer.

If you are familiar with Expat, you are probably wondering how the push interface provided by Expat was adapted to the pull API shown earlier. Expat allows us to suspend and resume parsing after every event and that's exactly what this implementation does. The performance cost of this constant suspension and resumption is about 35% of Expat's performance, which is not negligible but not the end of the world either.

All in, with all the name splitting and string constructions, parsing throughput on a 2010 Intel Core i7 laptop is about 37 MByte/sec, which should be sufficient for most applications.

While it is much easier to implement a conforming serializer from scratch, libstudxml reuses an existing and tested implementation in this case as well. It includes source code of a small C library for XML serialization called Genx (also MIT licensed) that was initially created by Tim Bray and significantly improved and extended over the past years as part of the XSD/e project.