The Template Implementation

The C++ template(1) facility, which effectively allows use of variables for types in declarations, is one of the newest features of the language.

GNU C++ is one of the first compilers to implement many of the template facilities currently defined by the ANSI committee.

Nevertheless, the template implementation is not yet complete. This chapter maps the current limitations of the GNU C++ template implementation.

• Template limitations: Limitations for function and class templates
• Function templates: Limitations for function templates
• Class templates: Limitations for class templates
• Template debugging: Debugging information for templates

Limitations for function and class templates

These limitations apply to any use of templates (function templates or class templates) with GNU C++:

Template definitions must be visible

When you compile code with templates, the template definitions must come first (before the compiler needs to expand them), and template definitions you use must be visible in the current scope.

Individual initializers needed for static data

Templates for static data in template classes do not work. See section Limitations for class templates.

Limitations for function templates

Function templates are implemented for the most part. The compiler can correctly determine template parameter values, and will delay instantiation of a function that uses templates until the requisite type information is available.

The following limitations remain:

• Narrowed specification: function declarations should not prevent template expansion. When you declare a function, GNU C++ interprets the declaration as an indication that you will provide a definition for that function. Therefore, GNU C++ does not use a template expansion if there is also an applicable declaration. GNU C++ only expands the template when there is no such declaration. The specification in Bjarne Stroustrup's The C++ Programming Language, Second Edition is narrower, and the GNU C++ implementation is now clearly incorrect. With this new specification, a declaration that corresponds to an instantiation of a function template only affects whether conversions are needed to use that version of the function. It should no longer prevent expansion of the template definition. For example, this code fragment must be treated differently:

  template <class X> X min (X& x1, X& x2) { ... }
  int min (int, int);
  ...
  int i; short s;
  min (i, s); // should call min(int,int)
              // derived from template
  ...
  

• The compiler does not yet understand function signatures where types are nested within template parameters. For example, a function like the following produces a syntax error on the closing `)' of the definition of the function f:

  template <class T> class A { public: T x; class Y {}; };
  template <class X> int f (A<X>::Y y) { ... }
  

• If you declare an inline function using templates, the compiler can only inline the code after the first time you use that function with whatever particular type signature the template was instantiated. Removing this limitation is akin to supporting nested function definitions in GNU C++; the limitation will probably remain until the more general problem of nested functions is solved.
• All the method templates (templates for member functions) for a class must be visible to the compiler when the class template is instantiated.

Limitations for class templates

Unfortunately, individual initializations of this sort are likely to be considered errors eventually; since they're needed now, you might want to flag places where you use them with comments to mark the need for a future transition.
• Member functions in template classes may not have results of nested type; GNU C++ signals a syntax error on the attempt. The following example illustrates this problem with an enum type alph:

  template <class T> class list {
    ...
    enum alph {a,b,c};
    alph bar();
    ...
  };
  
  template <class T>
  list<int>::alph list<int>::bar()  // Syntax error here
  {
  ...
  }
  

• A parsing bug makes it difficult to use preprocessor conditionals within templates. For example, in this code:

  template <class T>
  class list {
    ...
  #ifdef SYSWRONG
    T x;
  #endif
    ...
  }
  

  The preprocessor output leaves sourcefile line number information (lines like `# 6 "foo.cc"' when it expands the #ifdef block. These lines confuse the compiler while parsing templates, giving a syntax error. If you cannot avoid preprocessor conditionals in templates, you can suppress the line number information using the `-P' preprocessor option (but this will make debugging more difficult), by compiling the affected modules like this:

  g++ -P foo.cc -o foo
  

• Parsing errors are reported when templates are first instantiated---not on the template definition itself. In particular, if you do not instantiate a template definition at all, the compiler never reports any parsing errors that may be in the template definition.

Debugging information for templates

Debugging information for templates works for some object code formats, but not others. It works for stabs(2) (used primarily in A.OUT object code, but also in the Solaris 2 version of ELF), and the MIPS version of COFF debugging format.

DWARF support is currently minimal, and requires further development.