5. Extensions to the C++ Language

The GNU compiler provides these extensions to the C++ language (and you can also use most of the C language extensions in your C++ programs). If you want to write code that checks whether these features are available, you can test for the GNU compiler the same way as for C programs: check for a predefined macro __GNUC__. You can also use __GNUG__ to test specifically for GNU C++ (see section `Standard Predefined Macros' in The C Preprocessor).

5.1 Named Return Values in C++		Giving a name to C++ function return values.
5.2 Minimum and Maximum Operators in C++		C++ Minimum and maximum operators.
5.3 goto and Destructors in GNU C++		Goto is safe to use in C++ even when destructors are needed.
5.4 Declarations and Definitions in One Header		You can use a single C++ header file for both declarations and definitions.
5.5 Where's the Template?		Methods for ensuring that exactly one copy of each needed template instantiation is emitted.
5.6 Extracting the function pointer from a bound pointer to member function		You can extract a function pointer to the method denoted by a `->*' or `.*' expression.
5.7 Type Abstraction using Signatures		You can specify abstract types to get subtype polymorphism independent from inheritance.

                        

5.1 Named Return Values in C++

GNU C++ extends the function-definition syntax to allow you to specify a name for the result of a function outside the body of the definition, in C++ programs:

type functionname (args) return resultname; { ... body ... }

You can use this feature to avoid an extra constructor call when a function result has a class type. For example, consider a function m, declared as `X v = m ();', whose result is of class X:

X m () { X b; b.a = 23; return b; }

Although m appears to have no arguments, in fact it has one implicit argument: the address of the return value. At invocation, the address of enough space to hold v is sent in as the implicit argument. Then b is constructed and its a field is set to the value 23. Finally, a copy constructor (a constructor of the form `X(X&)') is applied to b, with the (implicit) return value location as the target, so that v is now bound to the return value.

But this is wasteful. The local b is declared just to hold something that will be copied right out. While a compiler that combined an "elision" algorithm with interprocedural data flow analysis could conceivably eliminate all of this, it is much more practical to allow you to assist the compiler in generating efficient code by manipulating the return value explicitly, thus avoiding the local variable and copy constructor altogether.

Using the extended GNU C++ function-definition syntax, you can avoid the temporary allocation and copying by naming r as your return value at the outset, and assigning to its a field directly:

X m () return r; { r.a = 23; }

The declaration of r is a standard, proper declaration, whose effects are executed before any of the body of m.

Functions of this type impose no additional restrictions; in particular, you can execute return statements, or return implicitly by reaching the end of the function body ("falling off the edge"). Cases like

X m () return r (23); { return; }

(or even `X m () return r (23); { }') are unambiguous, since the return value r has been initialized in either case. The following code may be hard to read, but also works predictably:

X m () return r; { X b; return b; }

The return value slot denoted by r is initialized at the outset, but the statement `return b;' overrides this value. The compiler deals with this by destroying r (calling the destructor if there is one, or doing nothing if there is not), and then reinitializing r with b.

This extension is provided primarily to help people who use overloaded operators, where there is a great need to control not just the arguments, but the return values of functions. For classes where the copy constructor incurs a heavy performance penalty (especially in the common case where there is a quick default constructor), this is a major savings. The disadvantage of this extension is that you do not control when the default constructor for the return value is called: it is always called at the beginning.

5.2 Minimum and Maximum Operators in C++

It is very convenient to have operators which return the "minimum" or the "maximum" of two arguments. In GNU C++ (but not in GNU C),

a <? b

is the minimum, returning the smaller of the numeric values a and b;

a >? b

is the maximum, returning the larger of the numeric values a and b.

These operations are not primitive in ordinary C++, since you can use a macro to return the minimum of two things in C++, as in the following example.

#define MIN(X,Y) ((X) < (Y) ? : (X) : (Y))

You might then use `int min = MIN (i, j);' to set min to the minimum value of variables i and j.

However, side effects in X or Y may cause unintended behavior. For example, MIN (i++, j++) will fail, incrementing the smaller counter twice. A GNU C extension allows you to write safe macros that avoid this kind of problem (see section Naming an Expression's Type). However, writing MIN and MAX as macros also forces you to use function-call notation for a fundamental arithmetic operation. Using GNU C++ extensions, you can write `int min = i <? j;' instead.

Since <? and >? are built into the compiler, they properly handle expressions with side-effects; `int min = i++ <? j++;' works correctly.

5.3 goto and Destructors in GNU C++

In C++ programs, you can safely use the goto statement. When you use it to exit a block which contains aggregates requiring destructors, the destructors will run before the goto transfers control.

The compiler still forbids using goto to enter a scope that requires constructors.

5.4 Declarations and Definitions in One Header

C++ object definitions can be quite complex. In principle, your source code will need two kinds of things for each object that you use across more than one source file. First, you need an interface specification, describing its structure with type declarations and function prototypes. Second, you need the implementation itself. It can be tedious to maintain a separate interface description in a header file, in parallel to the actual implementation. It is also dangerous, since separate interface and implementation definitions may not remain parallel.

With GNU C++, you can use a single header file for both purposes.

Warning: The mechanism to specify this is in transition. For the nonce, you must use one of two #pragma commands; in a future release of GNU C++, an alternative mechanism will make these #pragma commands unnecessary.

The header file contains the full definitions, but is marked with `#pragma interface' in the source code. This allows the compiler to use the header file only as an interface specification when ordinary source files incorporate it with #include. In the single source file where the full implementation belongs, you can use either a naming convention or `#pragma implementation' to indicate this alternate use of the header file.

#pragma interface

#pragma interface "subdir/objects.h"

Use this directive in header files that define object classes, to save space in most of the object files that use those classes. Normally, local copies of certain information (backup copies of inline member functions, debugging information, and the internal tables that implement virtual functions) must be kept in each object file that includes class definitions. You can use this pragma to avoid such duplication. When a header file containing `#pragma interface' is included in a compilation, this auxiliary information will not be generated (unless the main input source file itself uses `#pragma implementation'). Instead, the object files will contain references to be resolved at link time.

The second form of this directive is useful for the case where you have multiple headers with the same name in different directories. If you use this form, you must specify the same string to `#pragma implementation'.

#pragma implementation

#pragma implementation "objects.h"

Use this pragma in a main input file, when you want full output from included header files to be generated (and made globally visible). The included header file, in turn, should use `#pragma interface'. Backup copies of inline member functions, debugging information, and the internal tables used to implement virtual functions are all generated in implementation files.

If you use `#pragma implementation' with no argument, it applies to an include file with the same basename(2) as your source file. For example, in `allclass.cc', giving just `#pragma implementation' by itself is equivalent to `#pragma implementation "allclass.h"'.

In versions of GNU C++ prior to 2.6.0 `allclass.h' was treated as an implementation file whenever you would include it from `allclass.cc' even if you never specified `#pragma implementation'. This was deemed to be more trouble than it was worth, however, and disabled.

If you use an explicit `#pragma implementation', it must appear in your source file before you include the affected header files.

Use the string argument if you want a single implementation file to include code from multiple header files. (You must also use `#include' to include the header file; `#pragma implementation' only specifies how to use the file--it doesn't actually include it.)

There is no way to split up the contents of a single header file into multiple implementation files.

`#pragma implementation' and `#pragma interface' also have an effect on function inlining.

If you define a class in a header file marked with `#pragma interface', the effect on a function defined in that class is similar to an explicit extern declaration--the compiler emits no code at all to define an independent version of the function. Its definition is used only for inlining with its callers.

Conversely, when you include the same header file in a main source file that declares it as `#pragma implementation', the compiler emits code for the function itself; this defines a version of the function that can be found via pointers (or by callers compiled without inlining). If all calls to the function can be inlined, you can avoid emitting the function by compiling with `-fno-implement-inlines'. If any calls were not inlined, you will get linker errors.

5.5 Where's the Template?

C++ templates are the first language feature to require more intelligence from the environment than one usually finds on a UNIX system. Somehow the compiler and linker have to make sure that each template instance occurs exactly once in the executable if it is needed, and not at all otherwise. There are two basic approaches to this problem, which I will refer to as the Borland model and the Cfront model.

Borland model

Borland C++ solved the template instantiation problem by adding the code equivalent of common blocks to their linker; the compiler emits template instances in each translation unit that uses them, and the linker collapses them together. The advantage of this model is that the linker only has to consider the object files themselves; there is no external complexity to worry about. This disadvantage is that compilation time is increased because the template code is being compiled repeatedly. Code written for this model tends to include definitions of all templates in the header file, since they must be seen to be instantiated.

Cfront model

The AT&T C++ translator, Cfront, solved the template instantiation problem by creating the notion of a template repository, an automatically maintained place where template instances are stored. A more modern version of the repository works as follows: As individual object files are built, the compiler places any template definitions and instantiations encountered in the repository. At link time, the link wrapper adds in the objects in the repository and compiles any needed instances that were not previously emitted. The advantages of this model are more optimal compilation speed and the ability to use the system linker; to implement the Borland model a compiler vendor also needs to replace the linker. The disadvantages are vastly increased complexity, and thus potential for error; for some code this can be just as transparent, but in practice it can been very difficult to build multiple programs in one directory and one program in multiple directories. Code written for this model tends to separate definitions of non-inline member templates into a separate file, which should be compiled separately.

When used with GNU ld version 2.8 or later on an ELF system such as Linux/GNU or Solaris 2, or on Microsoft Windows, g++ supports the Borland model. On other systems, g++ implements neither automatic model.

A future version of g++ will support a hybrid model whereby the compiler will emit any instantiations for which the template definition is included in the compile, and store template definitions and instantiation context information into the object file for the rest. The link wrapper will extract that information as necessary and invoke the compiler to produce the remaining instantiations. The linker will then combine duplicate instantiations.

In the mean time, you have the following options for dealing with template instantiations:

1. Compile your template-using code with `-frepo'. The compiler will generate files with the extension `.rpo' listing all of the template instantiations used in the corresponding object files which could be instantiated there; the link wrapper, `collect2', will then update the `.rpo' files to tell the compiler where to place those instantiations and rebuild any affected object files. The link-time overhead is negligible after the first pass, as the compiler will continue to place the instantiations in the same files.

   This is your best option for application code written for the Borland model, as it will just work. Code written for the Cfront model will need to be modified so that the template definitions are available at one or more points of instantiation; usually this is as simple as adding #include <tmethods.cc> to the end of each template header.

   For library code, if you want the library to provide all of the template instantiations it needs, just try to link all of its object files together; the link will fail, but cause the instantiations to be generated as a side effect. Be warned, however, that this may cause conflicts if multiple libraries try to provide the same instantiations. For greater control, use explicit instantiation as described in the next option.

2. Compile your code with `-fno-implicit-templates' to disable the implicit generation of template instances, and explicitly instantiate all the ones you use. This approach requires more knowledge of exactly which instances you need than do the others, but it's less mysterious and allows greater control. You can scatter the explicit instantiations throughout your program, perhaps putting them in the translation units where the instances are used or the translation units that define the templates themselves; you can put all of the explicit instantiations you need into one big file; or you can create small files like

   #include "Foo.h" #include "Foo.cc" template class Foo<int>; template ostream& operator << (ostream&, const Foo<int>&);

   for each of the instances you need, and create a template instantiation library from those.

   If you are using Cfront-model code, you can probably get away with not using `-fno-implicit-templates' when compiling files that don't `#include' the member template definitions.

   If you use one big file to do the instantiations, you may want to compile it without `-fno-implicit-templates' so you get all of the instances required by your explicit instantiations (but not by any other files) without having to specify them as well.

   g++ has extended the template instantiation syntax outlined in the Working Paper to allow forward declaration of explicit instantiations and instantiation of the compiler support data for a template class (i.e. the vtable) without instantiating any of its members:

   extern template int max (int, int); inline template class Foo<int>;

3. Do nothing. Pretend g++ does implement automatic instantiation management. Code written for the Borland model will work fine, but each translation unit will contain instances of each of the templates it uses. In a large program, this can lead to an unacceptable amount of code duplication.

4. Add `#pragma interface' to all files containing template definitions. For each of these files, add `#pragma implementation "filename"' to the top of some `.C' file which `#include's it. Then compile everything with `-fexternal-templates'. The templates will then only be expanded in the translation unit which implements them (i.e. has a `#pragma implementation' line for the file where they live); all other files will use external references. If you're lucky, everything should work properly. If you get undefined symbol errors, you need to make sure that each template instance which is used in the program is used in the file which implements that template. If you don't have any use for a particular instance in that file, you can just instantiate it explicitly, using the syntax from the latest C++ working paper:

   template class A<int>; template ostream& operator << (ostream&, const A<int>&);

   This strategy will work with code written for either model. If you are using code written for the Cfront model, the file containing a class template and the file containing its member templates should be implemented in the same translation unit.

   A slight variation on this approach is to instead use the flag `-falt-external-templates'; this flag causes template instances to be emitted in the translation unit that implements the header where they are first instantiated, rather than the one which implements the file where the templates are defined. This header must be the same in all translation units, or things are likely to break.

   See section Declarations and Definitions in One Header, for more discussion of these pragmas.

5.6 Extracting the function pointer from a bound pointer to member function

In C++, pointer to member functions (PMFs) are implemented using a wide pointer of sorts to handle all the possible call mechanisms; the PMF needs to store information about how to adjust the `this' pointer, and if the function pointed to is virtual, where to find the vtable, and where in the vtable to look for the member function. If you are using PMFs in an inner loop, you should really reconsider that decision. If that is not an option, you can extract the pointer to the function that would be called for a given object/PMF pair and call it directly inside the inner loop, to save a bit of time.

Note that you will still be paying the penalty for the call through a function pointer; on most modern architectures, such a call defeats the branch prediction features of the CPU. This is also true of normal virtual function calls.

The syntax for this extension is

extern A a; extern int (A::*fp)(); typedef int (*fptr)(A *); fptr p = (fptr)(a.*fp);

You must specify `-Wno-pmf-conversions' to use this extension.

5.7 Type Abstraction using Signatures

In GNU C++, you can use the keyword signature to define a completely abstract class interface as a datatype. You can connect this abstraction with actual classes using signature pointers. If you want to use signatures, run the GNU compiler with the `-fhandle-signatures' command-line option. (With this option, the compiler reserves a second keyword sigof as well, for a future extension.)

Roughly, signatures are type abstractions or interfaces of classes. Some other languages have similar facilities. C++ signatures are related to ML's signatures, Haskell's type classes, definition modules in Modula-2, interface modules in Modula-3, abstract types in Emerald, type modules in Trellis/Owl, categories in Scratchpad II, and types in POOL-I. For a more detailed discussion of signatures, see Signatures: A Language Extension for Improving Type Abstraction and Subtype Polymorphism in C++ by Gerald Baumgartner and Vincent F. Russo (Tech report CSD--TR--95--051, Dept. of Computer Sciences, Purdue University, August 1995, a slightly improved version appeared in Software--Practice & Experience, 25(8), pp. 863--889, August 1995). You can get the tech report by anonymous FTP from ftp.cs.purdue.edu in `pub/gb/Signature-design.ps.gz'.

Syntactically, a signature declaration is a collection of member function declarations and nested type declarations. For example, this signature declaration defines a new abstract type S with member functions `int foo ()' and `int bar (int)':

signature S { int foo (); int bar (int); };

Since signature types do not include implementation definitions, you cannot write an instance of a signature directly. Instead, you can define a pointer to any class that contains the required interfaces as a signature pointer. Such a class implements the signature type.

To use a class as an implementation of S, you must ensure that the class has public member functions `int foo ()' and `int bar (int)'. The class can have other member functions as well, public or not; as long as it offers what's declared in the signature, it is suitable as an implementation of that signature type.

For example, suppose that C is a class that meets the requirements of signature S (C conforms to S). Then

C obj; S * p = &obj;

defines a signature pointer p and initializes it to point to an object of type C. The member function call `int i = p->foo ();' executes `obj.foo ()'.

Abstract virtual classes provide somewhat similar facilities in standard C++. There are two main advantages to using signatures instead:

1. Subtyping becomes independent from inheritance. A class or signature type T is a subtype of a signature type S independent of any inheritance hierarchy as long as all the member functions declared in S are also found in T. So you can define a subtype hierarchy that is completely independent from any inheritance (implementation) hierarchy, instead of being forced to use types that mirror the class inheritance hierarchy.

2. Signatures allow you to work with existing class hierarchies as implementations of a signature type. If those class hierarchies are only available in compiled form, you're out of luck with abstract virtual classes, since an abstract virtual class cannot be retrofitted on top of existing class hierarchies. So you would be required to write interface classes as subtypes of the abstract virtual class.

There is one more detail about signatures. A signature declaration can contain member function definitions as well as member function declarations. A signature member function with a full definition is called a default implementation; classes need not contain that particular interface in order to conform. For example, a class C can conform to the signature

signature T { int f (int); int f0 () { return f (0); }; };

whether or not C implements the member function `int f0 ()'. If you define C::f0, that definition takes precedence; otherwise, the default implementation S::f0 applies.