Skip to main content

[C++] Handling Exceptions in Constructors

When you use RAII idiom, there are often situations where constructors have to do complex tasks. These complex tasks can sometimes fail, resulting in throwing exceptions. This raises a concern: Is it okay to throw exceptions in constructors?

The first concern is memory leaks. Fortunately, memory leaks do not occur. Variables created on the stack are released through stack unwinding, and if an exception occurs during heap allocation with the new operator, the new operator automatically deallocates the memory and returns nullptr.

The next concern is whether the destructor of the member variables will be called correctly. However, this is also not a problem. When an exception occurs, member variables can be divided into three categories: fully initialized member variables, member variables being initialized, and uninitialized member variables. Fully initialized member variables have had their constructors called and memory allocations completed successfully. In the example code, this applies to the variable b. When an exception occurs, these variables have their destructors called and resources released properly.

Member variables being initialized, like c in the example code, are those in the middle of initialization when an exception occurs. If an exception occurs during member variable initialization, as in the code above, there is only one such variable. However, no member variables are initialized if an exception occurs in the constructor body.

Lastly, uninitialized member variables do not have their constructors called, so their destructors are not called either. However, this is correct behavior since their constructors have not been called yet.

What about inheritance? Will resources allocated by the parent class be released properly? Of course, there is no problem in this case either. When an exception occurs in a constructor, after the destructors of all fully initialized member variables have been called, the parent class's destructors are called to release resources. Thus, executing the code above will produce the following output.

Returning to the initial question: Is it okay to throw exceptions in constructors?

As seen above, there is no problem...... I hope to say. However, one assumption is required: If the class is well-designed, there is no problem with throwing exceptions in constructors.

So, what is a poorly designed class, and what problems can arise in such cases?

In the code above, the constructor of the class B allocates an A object on the heap and explicitly calls delete in the destructor to release it. If no exception occurs in the constructor, there is no problem. When B is allocated, and if an exception occurs, A is allocated, and when B is destroyed, A is destroyed as well.

However, if an exception occurs in B's constructor, a problem arises. When cleaning up B's member variables, the destructor of A is not called because a is merely a pointer to A. Moreover, since B has not been initialized, its destructor is not called, and a becomes a leaked memory. To prevent this, the lifetime of A should be matched to that of B, as the below code.


This article is the translation of my Korean post written in 2017. Looking back at it now, it is a good example of the importance of choosing meaningful variable names. I apologize for using example code with names like A, B, a, and b.

Comments

Popular posts from this blog

Iterator Adapters in Rust

An Iterator that takes another iterator and returns a new one is called an iterator adapter . The name "adapter" comes from one of the GoF's design patterns, the adapter pattern . However, in reality, it corresponds more to the decorator pattern , so if you pay too much attention to the name, you might get confused about its purpose. So it's better not to worry too much about the name. Enough complaining about the name, what does an iterator adapter do? An iterator adapter adds a task to be performed when the iterator iterates. This will be easier to understand when you see an example. The map function is one of the famous adapters. The iterator returned by the map function for those who have used functional languages iterates over new values transformed from the original values. Besides, various adapters are already implemented in the standard library. Among them, the most frequently used are those that are convenient to use with loops. Examples include the

What is the size of an empty object?

Consider a class like the one above. Commonly called an "empty class," this class has no internal variables. So, how big is this empty class? At first glance, the size should be 0 since there are no member variables. However, the size is never 0 in any language, whether Java, C#, C (in this case, a struct), or C++. This is to ensure that two different objects never have the same address. Empty classes typically have a size of 1 byte in a 32-bit environment and 2 bytes in a 64-bit environment. However, the exact size cannot be determined. According to the specification, the size just needs to be non-zero. The precise size depends on the implementation. This is a translation of my old Korean post written in 2015. Because the size can vary depending on the implementation, it is now possible to have different sizes (although still not 0). And Languages like Rust have even introduced zero-sized types . We will look at this topic in more detail at a future opportunity.