Static typing, pattern matching, and Inko's self-hosting compiler

Published on

It's been a while since our last Inko progress update. But worry not, we've been hard at work on the self-hosting Inko compiler.

The last progress report is from November 2019, almost seven months ago. The reason for the lack of updates is simple: we found the progress reports to be less useful than anticipated. Working towards a self-hosting compiler takes a long time, as past decisions may need to be re-evaluated as we go along. Such decisions include what syntax to use for certain features, how the type system should work, and more. This means that on a monthly basis there is less to talk about, resulting in progress reports being a bit boring.

In the last couple of weeks, quite some changes have been made; changes that we feel are worth discussing. We've also got some questions about the progress in general. With that in mind, we decided to give you all an update about what we have been up to since our last update.

Table of contents

If you would like to support the development of Inko, please donate to Inko on Open Collective or via GitHub Sponsors.

If you would like to engage with others interested in the development of Inko, please join our community on Matrix, Reddit, or follow the author of Inko on Twitter. For more information check out the Community page.

Windows CI now runs in VirtualBox

In the past we rented a VPS to run Windows tests on using GitLab CI. The costs of the VPS were quite high, and using Windows containers using Docker proved problematic. For example, updating the container alone could take hours; often just getting stuck for no clear reason.

To resolve these issues we changed our CI setup for Windows. We now run Windows tests in a VirtualBox VM, on a Mac Mini sponsored by MacStadium (which we were already using for running tests on macOS). This saves us just under €60 per month, and a lot of headaches.

Circular types are now supported

The Ruby compiler (and the code we have written so far for the self-hosting Inko compiler) now supports circular types, such as the following:

object A {
  @thing: B

  def init(thing: B) {
    @thing = thing
  }
}

object B {
  @thing: A

  def init(thing: A) {
    @thing = thing
  }
}

The compiler supports this by performing multiple passes over the AST when defining types, instead of only performing a single pass. This means that within the same module, it no longer matters what order types are defined in.

Now that circular types are supported, forward trait declarations are no longer necessary. This means the following is no longer valid:

trait A {}

object B {
  @thing: A

  def init(thing: A) {
    @thing = thing
  }
}

trait A {
  def foo
}

Nil has been split up

The type Nil has been split up in NilType and Nil, with Nil being a singleton instance of NilType; instead of Nil being both the object defined using the object keyword and a singleton instance. This change is made so that Nil becomes just another object instance, just like a String or Integer instance.

Modules are now first-class objects

Modules used to be emulated using objects defined in a "top-level" object. This object was never garbage-collected, and was used solely for storing module objects. This approach was a bit of a hack, and would complicate parts of the self-hosting compiler; without bringing any benefits.

To resolve this, the VM now has first-class support for modules. This is mostly an implementation detail, but makes it easier to maintain the VM and write a compiler for Inko.

Type parameters are no longer needed when re-opening objects

When re-opening a generic object, you no longer need to specify the names of the object's type parameters. Instead, they are implicitly made available. This means that you no longer need to write this:

object List!(T) {
  # ...
}

impl ToString for List!(T) {
  # ...
}

Instead, you can now write this:

object List!(T) {
  # ...
}

impl ToString for List {
  # ...
}

Boolean.not has been removed in favour of Boolean.false? and Boolean.true?

Using the method Boolean.not could lead to rather confusing code, such as the following:

def foo(value: Boolean) -> Boolean {
  value.not
}

Instead, you can now use Boolean.false? and Boolean.true?:

def foo(value: Boolean) -> Boolean {
  value.true?
}

The method Boolean.false? returns True if the receiver is False, while Boolean.true? does the opposite.

ByteArray is now in the prelude

The ByteArray type has been added to the prelude, removing the need for importing it manually using import std::byte_array::ByteArray.

Inko is now statically typed

Starting with the next release, Inko will be a statically typed language; instead of being gradually typed. When creating Inko we decided to go with gradual typing, with the hopes of it providing a bridge between the benefits of dynamic typing and static typing. For example, dynamically typed languages allow for rapid development and prototyping, while statically typed languages provide better safety guarantees.

As we continued work on the self-hosting compiler, two problems presented themselves:

  1. Supporting gradual typing complicates the compiler, and may prevent certain optimisations.
  2. Dynamically typed code does not play well with statically typed code.

An example of the first problem is keywords arguments. The VM has special knowledge of keyword arguments, allowing you to use them when the type a message is sent to is not known. This complicates the VM, adds overhead even when sending messages without keyword arguments, and is not used in Inko's standard library. To solve this, we will change how keyword arguments are implemented. This implementation requires that the compiler knows about all arguments available for a message. As such, it will not work when sending messages to dynamically typed values. As Inko's standard library makes little use of dynamic types, we felt this trade-off is worth it.

Certain optimisations may also be difficult to implement when supporting dynamic typing. For example, inlining is not possible when sending a message to a dynamic type; at least not at compile-time.

Gradual typing also doesn't work well if most of the code you interact with is statically typed, as is the case for Inko. This is because the compiler will forbid you from using a dynamic type (e.g. as an argument) when instead a static type (e.g. a String) is expected. To deal with this, every time you pass a dynamic type somewhere you will have to cast it to the expected type. Take this snippet for example:

import std::stdio::stdout

def say(message) {
  stdout.print(message)
}

say('Hello')

This code does not compile: the message argument is a dynamic type, but stdout.print expects a type that implements the ToString trait. To make this code compile, we would have to instead write the following:

import std::stdio::stdout
import std::conversion::ToString

def say(message) {
  stdout.print(message as ToString)
}

say('Hello')

This then brings the question: if we have to cast our types anyway, why not just use static typing? We wouldn't have to cast our types as much, and get compile-time safety.

With all this in mind, we have decided to make Inko a statically typed language.

Dynamic typing will be replaced with an Any trait, implemented by all objects. This trait exists for the odd case where you don't know what type you are dealing with at compile-time. Unlike a dynamic type, you can't send any messages to an Any as it does not respond to any messages; except for those available to the Object type. In other words, this compiles when using dynamic typing:

def example(message) {
  message.to_string
}

This does not compile when using the Any trait:

def example(message: Any) {
  message.to_string
}

This will not compile since neither Any nor Object (all types are an instance of Object) respond to the message to_string.

This change also introduces some changes to the syntax, and how method return types are inferred when left out. Method arguments must now specify either a default value, or a type. This is no longer valid syntax:

def example(message) {
  # ...
}

Leaving out the return type no longer results in it being inferred as a dynamic type. Instead, the return type is inferred as Nil and the method will always return Nil. This means that this:

def example {

}

Is now the same as this:

def example -> Nil {
  Nil
}

Inferring the return type as Nil (and having the method return Nil) makes it easier to write methods of which you want the return value to be ignored.

Support for pattern matching

Inko will support a limited form of pattern matching. Originally added to simplify the process of walking ASTs in the self-hosting compiler, there are other cases where having pattern matching can be useful. The syntax is inspired by that of Kotlin, and looks as follows:

def valid_number?(number: Integer) -> Boolean {
  match(number) {
    1..10 -> { True }
    else -> { False }
  }
}

Here we check if number falls in the range 1..10, returning True if this is the case. Pattern matching must be exhaustive, which we enforce by always requiring the presence of an else branch. Thus, the following is not valid:

match(number) {
  1..10 -> { True }
  50 -> { True }
}

You can also specify multiple patterns, and the case is matched if any of the patterns match:

def valid_number?(number: Integer) -> Boolean {
  match(number) {
    1, 2, 3, 4, 5 -> { True }
    5..8, 9..20 -> { True }
    else -> { False }
  }
}

Pattern matching expressions (as shown above) require that the patterns specified (1..10 for example) implement the trait std::operators::Match. This allows types to decide how and when they match a value. In case of the Range type, the implementation is as follows:

impl Match!(T) for Range {
  def =~(other: T) -> Boolean {
    cover?(other)
  }
}

We can also specify a "guard" in a pattern. If the pattern matches, the guard is evaluated. Only if both the pattern and guard return True do we consider the pattern as matched. For example:

def valid_token?(token: Token, current_line: Integer) -> Boolean {
  match(token.type) {
    'foo', 'bar', 'baz' when token.line == current_line -> { True }
    else -> { False }
  }
}

You can also bind the matched expression to a variable:

def valid_token?(token: Token, current_line: Integer) -> String {
  match(let type = token.type) {
    'foo', 'bar', 'baz' when token.line == current_line -> { type }
    else -> { 'unknown' }
  }
}

Pattern matching can also be used to perform a limited form of runtime type checking. When combined with a binding, the binding type is set to the matched type:

def visit(node: Node) {
  match(let matched = node) {
    as StringLiteral -> {
      # Here the type of "matched" is "StringLiteral"
    }
    as IntegerLiteral -> {
      # Here the type of "matched" is "IntegerLiteral"
    }
    else -> {
      # Here the type of "matched" is "Node"
    }
  }
}

The self-hosting compiler makes extensive use of this pattern when traversing the AST, removing the need for using the visitor pattern.

New Iterator methods

We've added several methods to the Iterator type: all?, zip, join, and reduce.

Iterator.all? is used to test if all values in an Iterator match a predicate:

Array.new(10, 20, 30).iter.all? do (value) { value.positive? } # => True

Iterator.zip is used to zip two iterators together:

let a = Array.new(10)
let b = Array.new(20)

a.iter.zip(b.iter).each do (pair) {
  pair.first  # => 10
  pair.second # => 20
}

Iterator.join is used to join the values in an Iterator together, producing a String:

Array.new(10, 20, 30).iter.join(',') # => '10,20,30'

Iterator.reduce is used to reduce an Iterator to a single value:

Array
  .new(1, 2, 3)
  .iter
  .reduce(0) do (total, current) { total + current } # => 6

Self-hosting compiler progress

Lots of progress has been made on Inko's self-hosting compiler. For the last several months we have focused on the type-checker, which is coming around nicely. A lot of Inko expressions can be type checked, though several important ones (e.g. sending messages to objects) are not yet supported.

With the compiler we're taking our time to make sure we don't make decisions we come to regret in the future. This slows down progress in the short term, but will save us time in the future. We hope to finish the self-hosting compiler by the end of 2020.

Improvements to the website

We have moved several pages on the website around, so they are in a more reasonable place. For example, the installation page is now located in the manual. The "Documentation" link at the top has been replaced with a "Learn" link that points straight to the manual, instead of pointing to a page telling users where to find documentation.

Plans for the coming months

In the coming months we will continue work on the type checker. Ideally we can also start working on designing the Intermediate Representation(s) of the compiler, used when optimising Inko code and generating bytecode. If you would like to stay up to date, please consider joining the growing community on Matrix.org or on Reddit.