Values and Classes

Storm, and therefore also Basic Storm, has three kinds of types: value types, class types, and actor types. This tutorial cover the first two of them: value types and class types. In particular, we will look at the differences between value types and class types, and how to implement new types in Basic Storm.

The code presented in this tutorial is available in the directory root/tutorials/types in the Storm release. You can run it by typing tutorials:types:main in the Basic Storm interactive top-loop.

Setup

Before starting to write code, we need somewhere to work. For this tutorial, it is enough to create a file with the extension .bs somewhere on your system. The tutorial uses the name types.bs, but any name that contains only letters will work. As a starting point, add the following code to the file:

void main() {
}

When you have created a file, open a terminal and change to the directory where you created the file. You can then run it by typing:

storm types.bs

Note that based on how you have installed Storm, you might need to modify the command-line slightly. Also, if you named your file something different than types.bs, you need to modify the name in the command accordingly.

If done correctly, Storm will find your file, notice that it contains a main function and execute it. Since the main function does not yet do anything, Storm will immediately exit without printing anything.

Value Types

Let's start by exploring value types. A value type is a type that has value semantics. This means that value types are copied when passed as a parameter to a function, returned from a function, and when using the assignment operator (=). Since value types are frequently copied, value types should typically be kept small in size, to avoid needlessly copying large amounts of data.

To illustrate the semantics in more detail, we start by creating a value type in the file we created earlier:

value Pair {
    Nat key;
    Str value;
}

So far the value type is very simple. It just contains two data members: a Nat named key and a Str named value. We also need to write some code to actually use the Pair. We do this in a function we call valueMain, so that we can compare the behavior to class types later on. As a start, we simply create an instance of the value and print its key and value:

void valueMain() {
    Pair p;
    print("Key: ${p.key}");
    print("Value: ${p.value}");
}

We also need to add the following to main to actually call our new function:

void main() {
    print("-- Value ---");
    valueMain();
}

If we run this program, we get the following output:

-- Value --
Key: 0
Value: 

Basic Storm requires that all variables are initialized to some well-defined value. In our example, Basic Storm helpfully generated a default constructor for our type. This constructory simply calls the default constructors of all data members in the type. This is why the key is guaranteed to contain the value 0, and why the value contains an empty string. In fact, since Str is a class type, Basic Storm had to heap-allocate an instance of the Str class for us, since null is not considered to be a valid value for variables by default.

Constructors

If we define any constructor for a type, Basic Storm will not generate a default constructor. This means that if we wish to require that users of our type should explicitly provide a key and a value to our class, we can simply define a constructor that accepts two parameters and initializes the member variables:

value Pair {
    Nat key;
    Str value;

    init(Nat key, Str value) {
        init {
            key = key;
            value = value;
        }
    }
}

As we can see, constructors are declared similarly to functions. However, the special name init (for initialize) is used, and no return type is specified. A constructor is also special in the sense that it may contain a init block. The init block in the constructor is where the object is actually initialized. As such it is not possible to access member variables and member functions before the init block, since the object does not exist at that point. The init block itself contains a list of initializers. Here, we specify that we wish to initialize the key member variable using the key variable, and the value member using the value variable. Note that while these initializers may look ambiguous at first glance, they are in fact not. The left hand side of the = sign must refer to a member variable. The right hand side is a general expression that might refer to either the parameter or the member variable. However, since the object has not yet been created, the parameter is the only option.

It is also possible to write the initializers as key(key) and value(value) respectively. This syntax allows passing multiple parameters to the constructor of the member variables if desired.

Since we now have created a constructor, we have now inhibited the creation of the default constructor. Since our constructor requires two parameters, the code we wrote that creates a Pair will fail to compile with the following error:

@/home/storm/types.bs(140-141): Syntax error:
No appropriate constructor for types.Pair found. Can not initialize 'p'.
Expected signature: __init(types.Pair&)

The error message tells us that it failed to find a suitable constructor to initialize the variable p. It expected to find a constructor that only accepts the implicit this parameter, but this constructor no longer exists. If we wish to allow default initialization, we can of course manually create a constructor that accepts no parameters as well:

init() {
    init {
        key = 1;
        value = "default";
    }
}

Note that we can omit the initialization of key and value if we wish. This makes Basic Storm initialize them using their default constructor as it did in the default constructor. It is also possible to give member variables default values in the declaration of the variable itself:

Nat key = 1;
Str value = "default";

Regardless of if we want to allow initializing a Pair without parameters or not, we can provide the parameters to the constructor in our valueMain function as follows:

Pair p(5, "test");

Creating a String Representation

One thing that is common to do is to make it convenient to print values. Let's see what happens by default by adding the following to our valueMain function:

print("Pair: ${p}");

In this case we get the following error message:

@/home/storm/types.bs(266-272): Syntax error:
Can not convert types.Pair& to string by calling 'toS'.

As we can see, the system complains that it does not know how to create a string representation of our Pair type. The best way to do this is by defining a toS member function that accepts a StrBuf. This makes Storm able to generate a toS function and a << operator for the type automatically. As such, we do this by defining a member function as follows:

void toS(StrBuf to) {
    to << "{ key: " << key << ", value: " << value << " }";
}

Note that we do not need an explicit return statement since the value of the last expression is returned automatically, and in this case it returns to exactly as we would like it to. After defining this function, we can run our program to get the following output:

-- Value --
Key: 5
Value: test
Pair: { key: 5, value: test }

Value Semantics

To illustrate the implications of the value semantics mentioned earlier, let's expand the code in our valueMain function a bit:

void valueMain() {
    Pair p(5, "test");
    print("Pair: ${p}");

    Pair copy = p;
    copy.key += 5;

    print("Copied: ${copy}");
    print("Original: ${p}");
}

In the new code, we first make a copy of p into a new variable using the = operator. This is equivalent to using the = operator to assign to existing variables. Then, the program adds 5 to key and prints the values of the two variables. Since the value semantics dictates that assignments create a copy, the modifications to copy are not visible in p:

-- Value --
Pair: { key: 5, value: test }
Copied: { key: 10, value: test }
Original: { key: 5, value: test }

Similarly, value types are copied when they are passed to functions. We can verify this by creating a function and calling it:

void valueFunction(Pair p) {
    p.key += 10;
    print("Inside the function: ${p}");
}

void valueMain() {
    Pair p(5, "test");
    print("Pair: ${p}");

    Pair copy = p;
    copy.key += 5;

    print("Copied: ${copy}");
    print("Original: ${p}");

    valueFunction(p);
    print("After calling the function: ${p}");
}

Again, even though the function modifies the parameter, the changes are not visible outside of the function since the Pair is copied when passed as a parameter to the function. As such, we get the following output as we would expect:

-- Value --
Pair: { key: 5, value: test }
Copied: { key: 10, value: test }
Original: { key: 5, value: test }
Inside the function: { key: 15, value: test }
After calling the function: { key: 5, value: test }

There is one exception to this behavior: the implicit this parameter of member functions. This exception exists to allow member functions in a value type to modify the contents of the type. For example, we can write a member function add to increase the value of key:

void add(Nat toAdd) {
    key += toAdd;
}

We can verify that it works as we would expect by replacing the use of the += operator with a call to the add function:

void valueFunction(Pair p) {
    p.add(10);
    print("Inside the function: ${p}");
}

void valueMain() {
    Pair p(5, "test");
    print("Pair: ${p}");

    Pair copy = p;
    copy.add(5);

    print("Copied: ${copy}");
    print("Original: ${p}");

    valueFunction(p);
    print("After calling the function: ${p}");
}

Class Types

Objects of class types are always allocated on the heap, and are managed by the garbage collector. As such, variables of a class type is just a pointer to memory on the heap. This means that class types have reference semantics. This means that assignments of class types simply make a copy of the pointer. The same is true when passing class types to and from functions. This means that for class types, any changes made to an object by a function are generally visible to the caller.

This does, of course, require some care. However, the fact that the system does not need to copy class types makes it cheap to pass them around, even if they are large objects. The pointer indirection also allows a pointer to a type T to actually refer to an object that is a subtype of T. This opens up for many common OOP techniques like inheritance, polymorphism, and dynamic dispatch.

To illustrate this, let's start by implementing a ClassPair class like the one for value types before. The only difference in the implementation so far is that we use the keyword class instead of value to define the type:

class ClassPair {
    Nat key;
    Str value;

    init(Nat key, Str value) {
        init {
            key = key;
            value = value;
        }
    }

    void add(Nat toAdd) {
        key += toAdd;
    }
}

To test our implementation, we define the functions classMain and classFunction as below:

void classFunction(ClassPair p) {
    p.add(10);
}

void classMain() {
    ClassPair p(5, "test");
    print("Pair: ${p}");

    ClassPair copy = p;
    copy.add(5);

    print("Copied: ${copy}");
    print("Original: ${p}");

    classFunction(p);
    print("After calling function: ${p}");
}

Finally, we call classMain from the main function so that the code is actually executed:

void main() {
    print("-- Value --");
    valueMain();

    print("-- Class --");
    classMain();
}

If we run the code now, we get the following output:

-- Class --
Pair: ClassPair @0x00007F891EF5A1D8
Copied: ClassPair @0x00007F891EF5A1D8
Original: ClassPair @0x00007F891EF5A1D8
After calling function: ClassPair @0x00007F891EF5A1D8

This is not too helpful. Since we have not defined a string representation yet, we get the default one provided by the Object class. This implementation simply prints the name of the class and its address in memory. We can, however, see one thing from the output: all printed objects are actually the same object since they have the same address. We can already at this point conclude that the behavior will be different from the version that uses value types.

To define the string representation for classes, we simply override the toS(StrBuf) function. Similarly to the situation for value types, this makes both the toS function and the << operator work as we would expect, since they both call the toS(StrBuf) function eventually. We do this by adding the following definition to the ClassPair class:

void toS(StrBuf to) : override {
    to << "{ key: " << key << ", value: " << value << " }";
}

After adding this definition, we get output that is easier to read:

-- Class --
Pair: { key: 5, value: test }
Copied: { key: 10, value: test }
Original: { key: 10, value: test }
After calling function: { key: 20, value: test }

From this output we can indeed see that the result is different from the value types. Since assignments and function calls have reference semantics, there is only ever one instance of the ClassPair type. Therefore, the modifications done through the copy variable and in the classFunction function are both visible through the original p variable. This is because all variables refer to the same piece of memory as we saw before.

We can verify this by comparing the identities of the objects in p and copy. For class types, the == operator is used to compare objects semantically (which allows using == for strings safely). The reason for this will become clear when we consider the semantics of class types in the presence of threads. Basic Storm does, however, provide the operator is to compare object identities. We can use it in the classMain function as follows:

void classMain() {
    ClassPair p(5, "test");
    print("Pair: ${p}");

    ClassPair copy = p;
    copy.add(5);

    print("Copied: ${copy}");
    print("Original: ${p}");
    if (p is copy) {
        print("p and copy are the same object");
    } else {
        print("p and copy are different objects");
    }

    classFunction(p);
    print("After calling function: ${p}");
}

As we would expect, the program prints the line p and copy are the same object.

Since classes have reference semantics by default, we need to explicitly make copies when we need them. Basic Storm still generates a copy constructor for this purpose. We just need to call it. We can do this by explicitly calling the constructor as follows:

ClassPair copy = ClassPair(p);

We can also use parentheses when declaring the variable, as this version always calls a constructor:

ClassPair copy(p);

Regardless of which option was selected, the program should now print p and copy are different objects, and the changes to the variable copied are no longer visible in the original:

-- Class --
Pair: { key: 5, value: test }
Copied: { key: 10, value: test }
Original: { key: 5, value: test }
p and copy are different objects
After calling function: { key: 15, value: test }

Note that the copy constructor only performs a shallow copy. That is, only the top-level object is copied, and any member variables that have reference semantics are still shared between the copies. In the case of the ClassPair type, the value member of the two instances will still refer to the same string since Str is a class type. This does, however, not matter here since strings in Storm are immutable. If we wish to create a deep copy, Storm provides a clone function that performs deep copies.

Polymorphism

One major benefit of class types having reference semantics is that the type of the value stored in a variable does not need to match the exact type of the variable. Storm, like many other languages, utilizes this to implement polymorphism with dynamic dispatch. To retain type safety, Storm allows any variable to contain a value of either the type of the variable, or a value of a subtype of the variable.

To illustrate this, let's define a subclass to the ClassPair class named CountedPair. The goal is to make the subclass keep track the number of times that key was changed. To make it possible to pass the CountedPair class to functions that accept a plain ClassPair value, we make CountedPair inherit from ClassPair:

class CountedPair extends ClassPair {
    Nat count;
}

To test our implementation, we also define a function that uses the subclass:

void inheritanceMain() {
    CountedPair p;
    classFunction(p);
    print("After calling the function: ${p}");
}

Finally, we update our main function as well:

void main() {
    print("-- Value --");
    valueMain();

    print("-- Class --");
    classMain();

    print("-- With inheritance --");
    inheritanceMain();
}

If we run the program at this point we will get the following error:

@/home/storm/types.bs(1257-1308): Syntax error:
No constructor (__init(types.ClassPair)) found in types.ClassPair.

The source reference points to the CountedPair class in its entirety. What the error tries to tell us is that the default constructor is unable to initialize the superclass ClassPair, since ClassPair itself does not have a default constructor. To solve this issue, we need to define a constructor in the CountedPair class that provides parameters to the superclass. In this case, it makes sense to make the constructor of the CountedPair class accept parameters and simply forward them to the superclass' constructor. We pass parameters to the superclass' constructor by providing them as "parameters" to the init block as follows:

init(Nat key, Str value) {
    init(key, value) {
        count = 0;
    }
}

As a sidenote, it is possible to split initialization into two steps by first initializing the superclass and then the derived class. This is typically only needed when two classes depend on each other. For completeness, it would look as follows:

init(Nat key, Str value) {
    // "this" is not available here.
    super(key, value);
    // "this" is available, and of type ClassPair.
    init {
        count = 0;
    }
    // "this" is available, and of type CountedPair.
}

Since the constructor of CountedPair now requires parameters, we also need to update the inheritanceMain function:

void inheritanceMain() {
    CountedPair p(5, "test");
    classFunction(p);
    print("After calling the function: ${p}");
}

At this point the program works, but it behaves exactly the same as if we would have used a ClassPair. Let's utilize the fact that we are using a derived class by overriding some member functions. We start by overriding toS to print the value of count:

void toS(StrBuf to) : override {
    super:toS(to);
    to << " (count = " << count << ")";
}

We maark the function with override to declare that we intend to override an existing function. This is not necessary for the program to work correctly. Specifying override, however, will cause Basic Storm to issue an error if the function does not override a function in a super class (e.g. because a mistyped parameter list), and can save quite a bit of effort in debugging the code.

The overridden toS function first calls the super class' implementation of the toS function (super:toS(to)) and then appends the value of the count value. Of course it is not necessary to call the super class' implementation if it is not necessary. After doing this we should see the value of count being outputted in addition to the previous output.

Finally, we override add as well to actually increase count whenever add is called. We can do this similarly to how we added an overriding function to the toS function:

void add(Nat toAdd) : override {
    count++;
    super:add(toAdd);
}

After this we can run the program once again. This time we can observe that our add function was called, and that it increased the value of count to 1. It is interesting to note that the correct version of the add function was called even though add was called by classFunction, which accepts a ClassPair as a parameter. Basic Storm implements this using vtable based dispatch, meaning that there is a (small) cost associated with overriding functions in this manner. In contrast to C++, Storm does not require annotating functions that should be possible to override in this manner with virtual. Rather, Storm determines which functions need dynamic dispatch automatically, and only incurs the cost when the feature is actually used.

Finally, it is worth noting that it is possible to use inheritance for value types as well. However, inheritance is typically not very commonly useful for value types. Since value types have by-value semantics, and are copied in assignments and function calls, a variable of a given type always contains a value of the specified type. As such, virtual dispatch does not work for value types.

Function Call Syntax

Basic Storm allows omitting empty parentheses when calling functions. As such, the call x.toS() is equivalent to x.toS. It is therefore the programmer's choice whether or not to add an empty pair of parentheses to remind the reader of the code that a function is called.

While this might seem strange at first, it can be useful when refactoring code as it allows seamlessly transitioning from member variables into getters and setters. This is perhaps most clearly illustrated with an example.

We previously created the class CountedPair with the goal of tracking changes to the key variable in the parent class. This works well if key is only ever changed by calling the add function. However, consider the following example:

void refactorMain() {
    CountedPair p(1, "test");
    p.key = 2 * p.key + 1;
    print("Count is now: ${p.count}");
}

Again, we also modify main to run the new function:

void main() {
    print("-- Value --");
    valueMain();

    print("-- Class --");
    classMain();

    print("-- With inheritance --");
    inheritanceMain();

    print("-- Refactor --");
    refactorMain();
}

If we run the program att this point, it prints that count is zero, even though we modified key. Of course this is because key is a public member variable that any part of the program can modify freely. As such, we are not currently able to modify the CountedPair class to update count whenever key is modified. For reasons like this, some languages encourage creating getter and setter functions preemptively in such cases. However, Basic Storm provides an alternative path forward. Since it is possible to call functions without parentheses, we can simply re-define key to be a member function that retrieves the value:

class ClassPair {
    private Nat myKey;
    Str value;

    init(Nat key, Str value) {
        init {
            myKey = key;
            value = value;
        }
    }

    Nat key() { myKey; }

    void add(Nat toAdd) {
        myKey += toAdd;
    }

    void toS(StrBuf to) {
        to << "{ key: " << myKey << ", value: " << value << " }";
    }
}

At this point we are almost done. However, if we run the program now, Basic Storm will point out a difference in the interface:

@/home/storm/types.bs(1743-1744): Syntax error:
Unable to assign a core.Nat to the value core.Nat. It is only possible to
assign to references (such as variables), and if an assignment function is available.

Basic Storm rightfully points out that it does not make sense to assign to the return value of a function (as we do in the refactorMain function). It does, however, suggest that an "assignment function" might solve our problem. An assignment function is a setter function that has been marked with the keyword assign. This makes Basic Storm consider them when a variable is being assigned to. In our case, we can define an assignment function for key as follows:

assign key(Nat newKey) {
    myKey = newKey;
}

Since the function is marked with assign, the function may be called instead of an assignment. In our case, the line:

p.key = 2 * p.key + 1;

Will be interpreted as:

p.key(2 * p.key() + 1);

This is exactly what we would have to do when using a getter and a setter. As such, if we run the program now, it works as expected except for the fact that count is still zero at the end of the program. Nowever, since we now have an assignment function, we can override it in the CountedPair class to get the desired behavior:

assign key(Nat newKey) : override {
    count++;
    super:key(newKey);
}

With this change, the program prints 1 as we would expect. Note, however, that we were able to perform this refactoring by only modifying the two classes. We did not have to change the code that used the classes at any point. In other words, we were able to "promote" the public member variable key into a setter and a getter function, without changing the interface of the class.

This approach can also be used to implement read-only variables: a read-only variable can simply be implemented as a private variable with only a get function.

Uniform Call Syntax

Another feature of Basic Storm that might appear strange at first is that Basic Storm implements what is sometimes called "uniform call syntax". This means that the two function calls a.fn(b) and f(a, b) are equivalent (except in cases where fn is both a member and a free function, then a.fn(b) prefers the member function while fn(a, b) prefers the free function).

This functionality makes it possible to extend the functionality of classes in other packages, while making the extensions appear as if they were present in the extended class from the beginning. To illustrate this, let's assume that we are writing a piece of code where we need to count how many times a particular character occurs in a string. We can not modify the Str class, so we simply define a free function to do the work for us:

Nat count(Str s, Char ch) {
    Nat result = 0;
    for (c in s) {
        if (c == ch)
            result++;
    }
    return result;
}

Since this is a free function, we can of course call it as:

Nat c = count("hello", 'l');
print("Count: ${c}");

However, due to uniform call syntax, and the fact that the first parameter is Str, we can also call the function as follows:

Nat c = "hello".count('l');
print("Count: ${c}");

While the change is fairly small, the second way of calling count looks as if count is a proper member of the Str class, even if our count function is not even in the same package as the Str class.

It is worth noting that it is possible to name the first parameter this as we have seen previously. This makes it possible to access members of the parameter more conveniently. For example, if we wished to create a reset function for our ClassPair class, we could implement it as follows:

void reset(ClassPair this) {
    key = 0;
}

Note that we do not need to write this.key = 0 to access the member of the class, since the parameter is named this. A parameter named this is, however, not special in any other ways. In particular, Storm does not perform dynamic dispatch on free functions, and value types are copied as usual.

The fact that Storm does not perform dynamic dispatch means that if we defined another reset function for CountedPair as follows:

void reset(CountedPair this) {
    key = 0;
    count = 0;
}

Then, the following code will still call the implementation that accepts a ClassPair and thereby not reset count. This is because the overload resolution only considers the static types of all parameters, and dynamic dispatch is not performed for non-members:

CountedPair counted(0, "");
ClassPair p = counted;
p.reset(); // Calls the "ClassPair" version.

Making reset into a member function will of course make the code work as we would expect, since Storm would use dynamic dispatch in that case.

The fact that value types are copied as usual for free functions means that we can not implement the reset function for the value-type Pair. For example:

void reset(Pair this) {
    key = 0;
}
Pair p(5, "");
p.reset();
print(p.key.toS()); // prints 5

This code would compile and run. However, since the variable p is copied into the parameter this of the function reset, the changes made inside reset are not visible to the caller, and therefore the print statement will still print 5.