Another language

06 Dec 2014 por asterite

Crystal has global type inference. You can program without type annotations except for a few cases where you are required to do so (generic type arguments).

This is a double-edged sword.

On one side, it’s really nice not to have to be explicit about types. This allows for very quick prototyping, similar to dynamic languages. You can quickly sketch an idea and evolve it without having to constantly retype things. This is also very helpful when refactoring and reorganizing code because the friction is very low. For example, when you extract a method out of a piece of code you just specify the names of the arguments and the compiler will take care of figuring out the arguments’ types and the return type. Or you can start using an instance variable right away by assigning some value to it, without having to declare it first with the possible types it will hold.

But there are also some downsides to this approach. Let’s analyze each of them.

Code becomes harder to understand and follow

Some say that without type annotations a code becomes harder to follow. Let’s look at an example.

def sum(values)
  count = 0
  values.each do |value|
    count += value
  end
  count
end

What is the type of values? How can one understand this piece of code without knowing what type it operates on?

When you learned how to program we are sure that at one point you were introduced to pseudocode. It looks similar to code you would find in real life, only it’s simplified: you can rarely see type annotations there. Type is kind of obvious in the context it is given. If you add type annotations and other information, it would make it harder to see the code’s intention. Sounds familiar?

In our view, Ruby code is very close to pseudocode. In the above code there are no type annotations. The algorithm is clear: iterate each of the items in values and add them to a count variable. That’s it. What’s the type of values? It doesn’t really matter. All we care about is that it can be iterated (with each) and that it can be summed. Additionally, then name sum, values and count help in understanding the method’s intention and the variables’ possible types.

Compare this to some other language where you would have to add some types:

interface Iterable<T>
  def each(&block : T ->)
end

interface Addable<T>
  def +(other : T)
end

def sum(values : Iterable<T>) where T : Addable<T>
  count = 0
  values.each do |value|
    count += value
  end
  count
end

Here we are telling the compiler that there is a type, Iterable that has an each method that yields elements of a generic type T. Then we also tell the compiler that there’s a type Addable<T> that has a method + that operates with values of its same type. Finally, we define the sum method to operate on values that belong to the Iterable<T> type, where each T implements Addable<T>.

To us, this last code is farther from the pseudocode we had in mind. Also, there’s more code to read and understand.

One possible counter-argument to this is that it now becomes easier to navigate the code. I want to know how each is implemented, or what Addable is about, and I know where to find that. We can’t do that with dynamic languages if we don’t have type annotations.

But… wait! Crystal is not a dynamic language. When it finishes compiling it assigns a type to every variable and method that were used. From that information we can recreate the code and make it browsable. And, in fact, Crystal has a (very experimental) tool for that, if you compile your code with crystal browser file.cr. This will open a web browser where you can see the type of variables and navigate methods. So in our opinion this is not a valid reason here.

This last thing also means that Crystal knows the type of a class’ instance variables. In Ruby you might be looking at some class with a @foo instance variable and wonder what its type is. Don’t worry, run crystal hierarchy file.cr and you will know the exact type. This will be mostly useful if you run it in your spec files because these show a class’ usage. And in the future it will be possible to view this information in a documentation format like RDoc.

Incremental compilation is not possible

Because there are no type annotations, the compiler needs to figure out the type of everything, each time, from scratch. There’s no way to compile a module to an object file, or some other format, and later reuse that information, because there isn’t any (type) information at first glance.

Luckily Crystal’s compiler is pretty fast: it takes between 5 and 10 seconds to compile the whole compiler (just 2.3 seconds of that time is spent in the type inference phase). For larger programs the time will become bigger, so we will have to find a solution for this problem.

Another issue is that it is hard to do a REPL. If code always had type annotations we could generate machine code and not worry about a variable’s type possibly changing, or even a type’s instance variables being created or changed.

Many times we were tempted to give up. “If we require the programmer to add type annotations in instance variables and methods, incremental compilation and a REPL might become a reality”. “At least the syntax would be pleasant”. Fourtunately, whenever one of us said that, another one replied with a big “NO”.

This “NO” has a very strong reason. If we change the language in that direction we will end up with another language. Crystal is a language where you can leave out type annotations (mind you: you can add type annotations if you really want to). But if you are forced to add them, then it will not be Crystal anymore. It will probably be very similar to one of the existing programming languages. And why would we want to do that? What would be the benefit of inventing yet another language that is similar to another language in existence (or maybe being developed)? None. It would be a waste of time, a duplication of effort.

True, if we don’t give up we will face harder challenges. Is incremental compilation really impossible? Couldn’t we think of a similar technique? Is a REPL doable somehow in this language? Very interesting questions arise. Challenging problems appear. Fun times come.

We believe a language where type annotations are not required from the compiler’s part, but where types are still there, is possible. We want a smart compiler. We don’t want to simplify the language to make our job, compiler writers, easier. We want to make programmers work easier. And fun :-)