Would you Like to Throw Exceptions from a Constructor?

Throwing an exception from a constructor has always been an arguable issue between my friends after a few beers of joy. It firstly came to me on an exam where question was asking me to design a class which lets not more than n number of instances to be constructed.

Not using the most straight-forward trick to have a private constructor, I decided to invoke an exception from constructor if there exists more than n instances of that particular class. Luckily, I was coding in Java and had the comfort of a managed environment under the belt. My answer wasn’t graded thankfully but this case became a major concern and I decided to make several experiments back in the old days.

According to what I have seen while debugging, Java simply assigns a number of nulls or zeros to all of the fields of the class unless constructor is successfully executed. Additionally, right after an exception is thrown from a constructor, a null reference is returned and the newly created invalid instance starts to wait for the garbage collector to pick it up and remove the useless object from memory permanently. In short terms, a new instance is created although platform tends to murder it quickly in silent.

What about Unmanaged Platforms?

It was nice to see somebody was taking care of such exceptional situations but what would you do if you were alone? Fault tolerance isn’t just catching all exceptions: (from slides of Bjarne Stroustrup)

1. The program must be in a valid state after an exception is caught
2. Resource leaks must be avoided

This will become a problem where memory management is all up to you. You will have to remove and cleanup resources after an exception has been thrown, otherwise it will cause a leakage in memory and you will have created a zombie object hanging around memory — attached to the mother thread.

Cleaning Up the Mess of C++

An object is not even an object once its constructor finished successfully. So, better try to release resources inside the constructor before you throw an exception. In this case, I’m always using a double-exception throwing pattern that I made up and passionately using it in similar situations I face.

Let’s ask the complier! Two lines of code is always more representative than 10 lines of words. Here below, I’ve written a class called “Foo”, not functional but dangerous and an exception “StreamInitException” was decided to be thrown if there happens a problem while longStreamChar is being initialized. Right after we allocate 10k bytes for longStreamChar, BOOM! Someone pushed the wrong button, somebody shut down the database or trouble makers’ spirits are near. Don’t try to catch it inside the constructor and see what happens. Many orphan kids, once we named them as longStreamChar, are all around and we can’t even access them.

#include <iostream>

struct StreamInitException {
    StreamInitException(char * m){ message = m; }
    char * message;
};

class Foo {
    public:
        Foo(){
            try {
                this->initCharStream();
            } catch(StreamInitException * ex) {
                this->~Foo(); //release resources
                throw ex; // exception originally thrown by initCharStream
            }
        }

        ~Foo(){
            if(this->longCharStream)
            delete this->longCharStream;
        }

        void initCharStream(){
            //create a long char array of 10k
            this->longCharStream = (char*)malloc(10000);
            throw new StreamInitException("Error occurred during init.");
        }
    private:
        char * longCharStream;
    };

int main(){

    Foo * foo = NULL;

    try{
        foo = new Foo;
    } catch(StreamInitException * ex){
        std::cout << ex->message << std::endl;
    }

    return 0;

}

Therefore, I’m handling init related bad moments of the object inside the constructor, deallocate memory etc. and re-throw the same exception finally after I get the hold of the control. Seems acceptable!

Consequently, I have a final question: do you prefer to throw an exception from constructors or search for other workarounds such as assigning the workload to a separate method called Init()?