Exception Handling in C++

This page might seem like it duplicates some of what we have just seen, but it is valuable because it gives a different perspective on the topic. Read chapter 1 on pages 15-60.

Overhead

When an exception is thrown, there's considerable runtime overhead (but it's good overhead, since objects are cleaned up automatically!). For this reason, you never want to use exceptions as part of your normal flow-of-control, no matter how tempting and clever it may seem. Exceptions should occur only rarely, so the overhead is piled on the exception and not on the normally executing code. One of the important design goals for exception handling was that it could be implemented with no impact on execution speed when it wasn't used; that is, as long as you don't throw an exception, your code runs as fast as it would without exception handling. Whether this is true depends on the particular compiler implementation you're using. (See the description of the "zero-cost model" later in this section).

You can think of a throw expression as a call to a special system function that takes the exception object as an argument and backtracks up the chain of execution. For this to work, extra information needs to be put on the stack by the compiler, to aid in stack unwinding. To understand this, you need to know about the runtime stack.

Whenever a function is called, information about that function is pushed onto the runtime stack in an activation record instance (ARI), also called a stack frame. A typical stack frame contains the address of the calling function (so execution can return to it), a pointer to the ARI of the function's static parent (the scope that lexically contains the called function, so variables global to the function can be accessed), and a pointer to the function that called it (its dynamic parent). The path that logically results from repetitively following the dynamic parent links is the dynamic chain, or call chain, that we've mentioned previously in this chapter. This is how execution can backtrack when an exception is thrown, and it is the mechanism that makes it possible for components developed without knowledge of one another to communicate errors at runtime.

To enable stack unwinding for exception handling, extra exception-related information about each function needs to be available for each stack frame. This information describes which destructors need to be called (so that local objects can be cleaned up), indicates whether the current function has a try block, and lists which exceptions the associated catch clauses can handle. There is space penalty for this extra information, so programs that support exception handling can be somewhat larger than those that don't. Even the compile-time size of programs using exception handling is greater, since the logic of how to generate the expanded stack frames during runtime must be generated by the compiler.

To illustrate this, we compiled the following program both with and without exception-handling support in Borland C++ Builder and Microsoft Visual C++:

//: C01:HasDestructor.cpp {O}
class HasDestructor {
public:
  ~HasDestructor() {}
};
 
void g(); // For all we know, g may throw.
 
void f() {
  HasDestructor h;
  g();
} ///:~

If exception handling is enabled, the compiler must keep information about ~HasDestructor( ) available at runtime in the ARI for f( ) (so it can destroy h properly should g( ) throw an exception). The following table summarizes the result of the compilations in terms of the size of the compiled (.obj) files (in bytes).

Compiler\Mode

With Exception Support

Without Exception Support

Borland

616

234

Microsoft

1162

680

Don't take the percentage differences between the two modes too seriously. Remember that exceptions (should) typically constitute a small part of a program, so the space overhead tends to be much smaller (usually between 5 and 15 percent).

This extra housekeeping slows down execution, but a clever compiler implementation avoids this. Since information about exception-handling code and the offsets of local objects can be computed once at compile time, such information can be kept in a single place associated with each function, but not in each ARI. You essentially remove exception overhead from each ARI and thus avoid the extra time to push them onto the stack. This approach is called the zero-cost model of exception handling, and the optimized storage mentioned earlier is known as the shadow stack.