10 - C H A P T E R -

C H A P T E R 10

Improving Program Performance

You can improve the performance of C++ functions by writing those functions in a manner that helps the compiler do a better job of optimizing them. Many books have been written on software performance in general and C++ in particular. For example, see C++ Programming Style by Tom Cargill (Addison-Wesley, 1992), Writing Efficient Programs by Jon Louis Bentley (Prentice-Hall, 1982), Efficient C++: Performance Programming Techniques by Dov Bulka and David Mayhew (Addison-Wesley, 2000), and Effective C++--50 Ways to Improve Your Programs and Designs, Second Edition, by Scott Meyers, (Addison-Wesley, 1998). This chapter does not repeat such valuable information, but discusses only those performance techniques that strongly affect the C++ compiler.

10.1 Avoiding Temporary Objects

C++ functions often produce implicit temporary objects, each of which must be created and destroyed. For non-trivial classes, the creation and destruction of temporary objects can be expensive in terms of processing time and memory usage. The C++ compiler does eliminate some temporary objects, but it cannot eliminate all of them.

Write functions to minimize the number of temporary objects as long as your programs remain comprehensible. Techniques include using explicit variables rather than implicit temporary objects and using reference parameters rather than value parameters. Another technique is to implement and use operations such as += rather than implementing and using only + and =. For example, the first line below introduces a temporary object for the result of a + b, while the second line does not.

T x = a + b;

T x( a ); x += b;

10.2 Using Inline Functions

Calls to small and quick functions can be smaller and quicker when expanded inline than when called normally. Conversely, calls to large or slow functions can be larger and slower when expanded inline than when branched to. Furthermore, all calls to an inline function must be recompiled whenever the function definition changes. Consequently, the decision to use inline functions requires considerable care.

Do not use inline functions when you anticipate changes to the function definition and recompiling all callers is expensive. Otherwise, use inline functions when the code to expand the function inline is smaller than the code to call the function or the application performs significantly faster with the function inline.

The compiler cannot inline all function calls, so making the most effective use of function inlining may require some source changes. Use the +w option to learn when function inlining does not occur. In the following situations, the compiler will not inline the function:

The function contains difficult control constructs, such as loops, switch statements, and try/catch statements. Many times these functions execute the difficult control constructs infrequently. To inline such a function, split the function into two parts, an inner part that contains the difficult control constructs and an outer part that decides whether or not to call the inner part. This technique of separating the infrequent part from the frequent part of a function can improve performance even when the compiler can inline the full function.

The inline function body is large or complicated. Apparently simple function bodies may be complicated because of calls to other inline functions within the body, or because of implicit constructor and destructor calls (as often occurs in constructors and destructors for derived classes). For such functions, inline expansion rarely provides significant performance improvement, and the function is best left uninlined.

The arguments to an inline function call are large or complicated. The compiler is particularly sensitive when the object for an inline member function call is itself the result of an inline function call. To inline functions with complicated arguments, simply compute the function arguments into local variables and then pass the variables to the function.

10.3 Using Default Operators

If a class definition does not declare a parameterless constructor, a copy constructor, a copy assignment operator, or a destructor, the compiler will implicitly declare them. These are called default operators. A C-like struct has these default operators. When the compiler builds a default operator, it knows a great deal about the work that needs to be done and can produce very good code. This code is often much faster than user-written code because the compiler can take advantage of assembly-level facilities while the programmer usually cannot. So, when the default operators do what is needed, the program should not declare user-defined versions of these operators.

Default operators are inline functions, so do not use default operators when inline functions are inappropriate (see the previous section). Otherwise, default operators are appropriate when:

The user-written parameterless constructor would only call parameterless constructors for its base objects and member variables. Primitive types effectively have "do nothing" parameterless constructors.

The user-written copy constructor would simply copy all base objects and member variables.

The user-written copy assignment operator would simply copy all base objects and member variables.

The user-written destructor would be empty.

Some C++ programming texts suggest that class programmers always define all operators so that any reader of the code will know that the class programmer did not forget to consider the semantics of the default operators. Obviously, this advice interferes with the optimization discussed above. The resolution of the conflict is to place a comment in the code stating that the class is using the default operator.

10.4 Using Value Classes

C++ classes, including structures and unions, are passed and returned by value. For Plain-Old-Data (POD) classes, the C++ compiler is required to pass the struct as would the C compiler. Objects of these classes are passed directly. For objects of classes with user-defined copy constructors, the compiler is effectively required to construct a copy of the object, pass a pointer to the copy, and destruct the copy after the return. Objects of these classes are passed indirectly. For classes that fall between these two requirements, the compiler can choose. However, this choice affects binary compatibility, so the compiler must choose consistently for every class.

For most compilers, passing objects directly can result in faster execution. This execution improvement is particularly noticeable with small value classes, such as complex numbers or probability values. You can sometimes improve program efficiency by designing classes that are more likely to be passed directly than indirectly.

In compatibility mode (-compat[=4]), a class is passed indirectly if it has any one of the following:

A user-defined constructor

A virtual function

A virtual base class

A base that is passed indirectly

A non-static data member that is passed indirectly

Otherwise, the class is passed directly.

In standard mode (the default mode), a class is passed indirectly if it has any one of the following:

A user-defined copy constructor

A user-defined destructor

A base that is passed indirectly

A non-static data member that is passed indirectly

Otherwise, the class is passed directly.

10.4.1 Choosing to Pass Classes Directly

To maximize the chance that a class will be passed directly:

Use default constructors, especially the default copy constructor, where possible.

Use the default destructor where possible. The default destructor is not virtual, therefore a class with a default destructor should generally not be a base class.

Avoid virtual functions and virtual bases.

10.4.2 Passing Classes Directly on Various Processors

Classes (and unions) that are passed directly by the C++ compiler are passed exactly as the C compiler would pass a struct (or union). However, C++ structs and unions are passed differently on different architectures.


Architecture	Description
SPARC V7/V8	Structs and unions are passed and returned by allocating storage within the caller and passing a pointer to that storage. (That is, all structs and unions are passed by reference.)
SPARC V9	Structs with a size no greater than 16 bytes (32 bytes) are passed (returned) in registers. Unions and all other structs are passed and returned by allocating storage within the caller and passing a pointer to that storage. (That is, small structs are passed in registers; unions and large structs are passed by reference.) As a consequence, small value classes are passed as efficiently as primitive types.
IA platforms	Structs and unions are passed by allocating space on the stack and copying the argument onto the stack. Structs and unions are returned by allocating a temporary object in the caller's frame and passing the address of the temporary object as an implicit first parameter.

10.5 Cache Member Variables

Accessing member variables is a common operation in C++ member functions.

The compiler must often load member variables from memory through the this pointer. Because values are being loaded through a pointer, the compiler sometimes cannot determine when a second load must be performed or whether the value loaded before is still valid. In these cases, the compiler must choose the safe, but slow, approach and reload the member variable each time it is accessed.

You can avoid unnecessary memory reloads by explicitly caching the values of member variables in local variables, as follows:

Declare a local variable and initialize it with the value of the member variable.

Use the local variable in place of the member variable throughout the function.

If the local variable changes, assign the final value of the local variable to the member variable. However, this optimization may yield undesired results if the member function calls another member function on that object.

This optimization is most productive when the values can reside in registers, as is the case with primitive types. The optimization may also be productive for memory-based values because the reduced aliasing gives the compiler more opportunity to optimize.

This optimization may be counter-productive if the member variable is often passed by reference, either explicitly or implicitly.

On occasion, the desired semantics of a class requires explicit caching of member variables, for instance when there is a potential alias between the current object and one of the member function's arguments. For example:

complex& operator*= (complex& left, complex& right)

  left.real = left.real * right.real + left.imag * right.imag;

  left.imag = left.real * right.imag + left.image * right.real;

will yield unintended results when called with:

x*=x;