XML: C++ Introspection by Extensible Visitation

[Extended Abstract] DRAFT NOT FOR PUBLICATION

Kurt Stephens

ION, Inc.

May 22, 2001

Abstract

Many problems with object serialization, object conversion, debugging, and object inspectors can be solved using meta-object protocols (MOP). The C++ language lacks a formal MOP, although some are available as source pre-processors, debugger information, or as a compiler extension. A full MOP is not necessary for many classes of problems where basic introspection is useful. This paper describes a portable and lightweight technique: extensible visitation of objects as an introspection primitive.

Table of Contents

1. Introduction
2. Visitation as an Introspection Primitive
3. XVF
Visitor classes
Visitee type objects
Visitor/Visitee Interface
Visitation Methods
Visitor Messages
4. Handling const-ness
5. Handling visit-by-value
6. Visitation Attributes
7. Examples
A Structured Data Dumper
An Archiver
XML I/O
Dynamic GUI generation
Reflective Accessors
XDR Bridge
8. Related Work
9. Future Work
10. Conclusion
11. References

Chapter 1. Introduction

Access to meta-type information about an application framework's classes is valuable. Some approaches use aspect-oriented techniques [OPENCXX], pre-compilation [OPENC++, IGUANA], debugger information [???] and compiler extension [GCC]. Extensible visitation, based on the Visitor design pattern [GAMMA95] is a lightweight and portable mechanism, similar to XDR [XDR] for interfacing different introspections with little impact to an existing C++ framework.

Chapter 2. Visitation as an Introspection Primitive

Standard C++ ostream classes are visitors that produce output as a side-effect of visiting typed data. Developers must create ostream visitation methods for each new application class. Most visitation methods added to C++ classes are specialized for different visitors, yet do mostly the same thing; recursively visit all data members.

For example:

class Foo
{
 private:
  double x, y; // data members
  int i3[3];
 public:
  Foo() { ... }
  Foo(double _x, double _y) { ... }

  // Visitation Methods, one for each type of visitor.

 // ostream dump method.
 ostream &print(ostream &os)
 {
   os << "Foo(" << x ", " << y << ", ";
   os << "{ "
      << i3[0] << ", "
      << i3[1] << ", "
      << i3[2]
      << " }";
   return os << ")";
 }

 // XML output method.
 ostream &print_xml(ostream &os)
 {
   os << "<Foo>";
   os << "<x>"; print_xml(x, os); os << "</x>";
   os << "<y>"; print_xml(y, os); os << "</y>";
   os << "<i3>";
   print_xml(i3[0], os);
   print_xml(i3[1], os);
   print_xml(i3[2], os);
   os << "</i3>";
   return os << "</Foo>";
 }

 // inspector generation method.
 inspector &generate_inspector(inspector &is)
 {
   is.begin_frame("Foo", this);
   is.generate_inspector("x", &x);
   is.generate_inspector("y", &y);
   is.begin_frame("i3[]", &i3);
   is.generate_inspector("0", &i3[0]);
   is.generate_inspector("1", &i3[1]);
   is.generate_inspector("2", &i3[2]);
   is.end_frame(&i3);
   is.end_frame(this);
   return is;
 }

 // XDR support method
 int xdr(XDR *xdr)
 {
   return
   xdr_double(xdr, &x) &&
   xdr_double(xdr, &y) &&
   xdr_int(xdr, &i3[0]) &&
   xdr_int(xdr, &i3[1]) &&
   xdr_int(xdr, &i3[2]);
 }
 // ... other visitor types
}

In the example above, more visitation methods are needed as the application classes (visitees) interface with more visitors. Adding, removing or renaming data members becomes a maintenance problem, since each visitation method must be updated. Likewise, adding new visitations becomes difficult as the application grows, since each application class must be updated.

If the number of application classes in a system is N and the number of visitor classes is M, the complexity of the application-visitor interfaces will be N * M, if visitor-specific methods are implemented for each application class. If a single generic visitation method is used for each visitor class, the complexity of the application-visitor interfaces will be N + M.

Ideally, a single visitation method suffices for most types of visitation. Essentially we want to move the complexity from the visited classes to the visitors, since the number of visitor classes will be small compared to the visited classes. Visitor classes tend to interface auxiliary systems and protocols, such as XML and other I/O systems, and change less frequently than the application classes.

Chapter 3. XVF

XVF (eXtensible Visitation Framework) is a C++ template-based framework for developing visitors and generic visitation methods. In XVF, a generic, introspection visitor is messaged with type, naming and layout information about the visited objects. The visitor uses the introspective messages in any useful manner and controls the recursion during visitation. The visitor's introspection methods are virtual, such that any visitor can be used in any visitation traversal. The set of introspection methods in the visitor closely follows a decomposition of the data types in C++: structures (classes), pointers, arrays and primitive datums (char, int, double).

Visitor classes

Visitors are sub-classes of XVFVisitor. XVFVisitor has methods for keeping track of what object is being visited, what members are being visited. Each intrinsic type (e.g. char, int, double) has a visitation method specialized for it in the visitor class.

Visitors are messaged with information about the visitee during visitation traversal.

Visitee type objects

The XVFType class provides type object information about types (e.g. type name and size), allocation methods, and assignment.

XVFType::type("Foo") returns the type object for the Foo class.

XVFType objects are constructed dynamically by class templates. There are five XVFType templates: intrinsic types (i.e. char, int, unsigned long), abstract classes crea(i.e. non-instantiable classes), concrete classes (i.e. instantiable classes), pointers, and arrays.

The XVFTypeIntrinsic<T> template creates XVFType classes for intrinsic types. The templates XVFTypeClass<T> and XVFTypeAbstract<T> create type objects for instantiable and abstract classes, respectively. Type objects for intrinsic types are created "by hand" in XVF. Pointer and array type objects are constructed by XVFTypePtr<T> and XVFTypeArray<T>, respectively. XVFType templates are never instatiated by the user of XVF, but are created automatically by the instatiation of xvf_type(T*) template functions based on pointers to type T.

Each type object implements a virtual visit(void *data, XVFVisitor *V) method that is specialized by the type object template to call xvf_visit((T *) data, V);

Visitor/Visitee Interface

Visitee classes are not sub-classes of a special "Visitee" class. They are "glued" to the XVF library by static type polymorphic functions:

  • XVFType *xvf_type(T *x)

    returns the XVFType object for type T.

  • T *xvf_alloc(T *dummy)

    allocates an object of type T.

  • void xvf_dealloc(T *x)

    deallocates an object of type T.

  • void xvf_visit(T *x, XVFVisitor *V, int opts)

    recursively visit an object of type T with visitor V.

Note: XVF does not dispatch using const &T because then void could not be handled in a unified manner and because datums are visited by address.

Visitation Methods

Any type T that implements a xvf_type(T *x) and xvf_visit(T *x, XVFVisitor *V, int opts) (the opts argument is explained later) functions can be visited by any XVFVisitor object.

XVF supplies macros for constructing a class' visitation method, visitation and type object accessor functions.

For example:

// Declare xvf_type(Foo *)
XVF_CLASS_BEGIN(Foo);

class Foo {
  double x, y;
  int i3[3];
public:
  Foo() { ... }
  Foo(double x, double y) { ... }

  // xvf_type() support
  XVF_CLASS_METHODS(XVF_);

  // xvf_visit() support
  XVF_CLASS_VISIT_BEGIN_INLINE(Foo)
  {
     XVF_MEMB(x);
     XVF_MEMB(y);
     XVF_MEMB(i3);
  }
  XVF_CLASS_VISIT_END()
};

// Define xvf_visit(Foo *)
XVF_CLASS_END(Foo);

// Define xvf_type(Foo *)
XVF_CLASS_INSTANCE(Foo);

Visitor Messages

The visitor is messaged with information about the visited object during traversal. The visitor can dynamically control recursion for any visitation member. The XVF_MEMB() macro above expands into messages to the visitor.

For example: any visitor V visiting Foo F will be have the following methods called on it:

xvf_visit(&F, V);

is equivalent to:

V->this_begin(xvf_type(&F), &F, opts);
 if ( V->memb(xvf_type(&F.x), "x", &F.x, opts) ) {
   V->memb_begin(xvf_type(&F.x), "x", &F.x, opts);
    xvf_visit(&F.x, V, opts);
   V->memb_end(xvf_type(&F.x), "x", &F.x, opts);
 }
 if ( V->memb(xvf_type(&F.y), "y", &F.y, opts) ) {
   V->memb_begin(xvf_type(&F.y), "y", &F.y, opts);
    xvf_visit(&F.y, V, opts);
   V->memb_end(xvf_type(&F.y), "y" &F.y, opts);
 }
 if ( V->memb(xvf_type(&F.i3), "i3", &F.y, opts) ) {
   V->memb_begin(xvf_type(&F.i3), "i3", &F.y, opts);
     V->arry_begin(xvf_type(&F.i3),
                   xvf_type(&F.i3[0]),
                   &F.i3, 3, opts);
       V->elem_begin(0, opts);
         xvf_visit(&F.i3[0], V, opts);
       V->elem_end(0, opts);
       V->elem_begin(1, opts);
         xvf_visit(&F.i3[1], V, opts);
       V->elem_end(1, opts);
       V->elem_begin(2, opts);
         xvf_visit(&F.i3[2], V, opts);
       V->elem_end(3, opts);
     V->arry_end(xvf_type(&F.i3),
                 xvf_type(&F.i3[0]),
                 &F.i3, 3, opts);
   V->memb_end(xvf_type(&F.i3), "i3" &F.y, opts);
 }
V->this_end();

Chapter 4. Handling const-ness

The opts argument to the visit functions relays const-ness to the visitor without requiring different routines and type objects for const and non-const visitation.

If opts == XVF_NON_CONST in V->visit(T *data, int opts), V->visit() has license to modify *data.

For example: an inspector visitor may generate an editable or non-editable field depending if opts == XVF_NON_CONST or opts == XVF_CONST.

Chapter 5. Handling visit-by-value

Sometimes the visitation semantics of a class' members should been through a pair of accessor methods or functions, rather than directly on an object data member.

For example:

XVF_CLASS_BEGIN(Temperature);
class Temperature {
  private:
    double _f; // Fahrenheit: f = (c * 1.8) + 32
    double _c; // Celsius:    c = (f - 32) / 1.8

  public:
  Temperature() { celsius(0); }

  // Getter, setter.
  double fahrenheit() const { return _f; }
  void fahrenheit(double x) {
    _f = x; _c = (_f - 32) / 1.8;
   }

  // Getter, setter.
  double celsius() const { return _c; }
  void celsius(double x) {
    _c = x; _f = _c * 1.8 + 32;
  }

  XVF_CLASS_METHODS(XVF_);
  XVF_CLASS_VISIT_BEGIN_INLINE(Temperature)
  {
     // Visit by Value
     XVF_VAL("fahrenheit", // name
             double,       // type
             _xvf_this->fahrenheit(),
                           // getter
             _xvfv_this->fahrenheit(XVF_TEMP)
	                   // setter
     );
     XVF_VAL("celsius",
             double,
             _xvfv_this->celsius(),
             _xvfv_this->celsius(XVF_TEMP));
  }
  XVF_CLASS_VISIT_END()
};
XVF_CLASS_END(Temperature);
XVF_CLASS_INSTANCE(Temperature);

If any visitor updates to the "fahrenheit" or "celsius" members of a Temperature object both the Fahrenheit and Celsius values will be synchronized through the getter and setter methods in Temperature.

Chapter 6. Visitation Attributes

A visitation method can relay visitor-specific information using visitation attributes.

Imagine a visitor that needs to know how each member is protected in the C++ class:

...
class Bar {
private:
  int a; // a is private.
protected:
  float b; // a is protected.
public:
  char *c; // c is public.

  Bar(...)

  XVF_CLASS_VISIT_BEGIN_INLINE(Bar)
  {
    // Save and set attribute value.
    XVF_ATTR_BEGIN("C++:protection", "private");
    XVF_MEMB(a);

    // Set attribute value.
    XVF_ATTR("C++:protection", "protected");
    XVF_MEMB(b);

    XVF_ATTR("C++:protection", "public");
    XVF_MEMB(c);

    // Restore attribute value.
    XVF_ATTR_END("C++:protection");
  }
  XVF_CLASS_VISIT_END();
};
...

The specialized visitor will check the attribute named "C++:protection" for its value and produce results accordingly.

Chapter 7. Examples

Here is a sample use of different visitor objects using the same visit methods for class Foo. The visitor code is not presented here.

A Structured Data Dumper

A structured data visitor writes a human-readable representation to an ostream.

XVF_Dump_Out<ostream> xvf(cout);
Foo p(1, 2);

xvfv << p;

An Archiver

An archiver visitor writes a machine-readable representation to an ostream.

XVF_Archive_Out<ostream> xvfv(cout);
Foo p(1, 2);

xvfv << p;

XML I/O

An XML output visitor writes objects as an XML stream.

XVF_XML_Out<ostream> xvfv(cout);
Foo p(1, 2);

xvfv << p;

An XML input visitor reads an XML stream into a object.

XVFV_XML_In<istream> xvfv(cin);
Foo p;

xvfv >> p;

Dynamic GUI generation

An inspector visitor generates TCL/TK code to create an inspector GUI.

Tcl_Interp *interp;
XVF_TK_Inspector xvfv(interp);
Foo p(1, 2);

xvfv << p;
TK widget generated by XVF_TK_Inspector.

Reflective Accessors

XVF_Getter xvfg;
Foo p(1, 2);
double py;

xvfg(&p, xvf_type(&p), "y", &py); // py = p.y

XVF_Setter xvfs;
xvfs(&p, xvf_type(&p), "x", &py); // p.x = py

XDR Bridge

XDR *xdr;
XVF_XDR xvfv(xdr);

xvfv << p;

Chapter 8. Related Work

EXternal Data Representation [XDR] uses generic visitation for encoding, decoding and freeing C objects. XDR is used primarily for Remote Procedure Calls. XDR visitors cannot be sub-classed directly and do not provide any mechanisms for XDR visitors to determine member names.

OpenC++ provides introspection using pre-compiler techniques. Iguana [IGUANA] seems to be based on pre-compiler techniques as well, but no public implementations are offered.

Chapter 9. Future Work

Namespaces, enumerations, bit-fields, unions, function pointers and methods are not yet handled. Class versioning for backward compatibility is not currently handled. A pre-processor to parse class definitions for members and insert xvf_visit() functions automatically would be helpful; OpenC++ might be useful here. A C interface for legacy code.

Chapter 10. Conclusion

This paper describes a a method for lightweight and portable introspection in C++ using visitation. An prototype implementation of XVF is available at URL: http://www.ionink.com/research.

Chapter 11. References