Switch branches/tags
Nothing to show
Find file Copy path
Fetching contributors…
Cannot retrieve contributors at this time
410 lines (310 sloc) 16.7 KB

Table of Contents generated with DocToc

Platform Design

Note: This document is a work in progress.

Siren is a programming platform around the idea of live objects, just like other Smalltalk dialects (Smalltalk, Self, etc.). In this kind of system, programs are written in terms of interactions between objects, a style of programming called Object Oriented. But while most of the influence in Siren comes from its close cousin Self, there's also a lot of design decisions that are informed by functional programming (in languages like ML and Haskell) and by Lisp dialects. This document describes the approach to design that's used in the core of Siren's platform.


Siren is a pure object oriented language using prototypes but being limited to single delegation. Objects can be extended arbitrarily and extensions are restricted to a lexical (message) region. Messages are stored in objects as unique values, rather than Strings, which means that collisions are not possible when extending an object. Names that refer to a message are resolved lexically, so different lexical regions may have a particular message name (e.g.: as-string) resolve to two different message objects in the same object.

Since one of its goals is to provide an interactive platform, Siren is a fully reflective system, and meta-data can be attached to any object, as well as queried at any time during the execution of the program. Users need to acquire the proper Mirror to reflect on the parts of the object they're interested in.

Another major goal in Siren is to be used as a programming language for teaching. Within this goal, the ability of understanding programs through examples, exploring them and extrapolating from them is essential. This fits nicely with the prototypical OO model, since the core idea in the model is to create objects that exemplify a concept (so they're fully usable on their own) and extrapolate from them.

Furthermore, to make reasoning about programs simpler, and live programming feasible, most of Siren's functions have no side-effects, and the ones that do are explicitly marked (with a ! suffix in the message name).

Design Goals

  • Correctness: The system should guide people to writing correct programs by design. This can be done by modeling the world such that objects can't enter invalid states.

  • Exploration: The system should support exploratory programming. Users should be able to experiment with their programs without the risk of unknowingly launching missiles.

  • Understandability: The objects should be easy to understand, which tends to mean they shouldn't do too much, and they should be aptly named after what they do.

  • Discoverability: The system should be discoverable, for this objects should be tagged with enough meta-information such that the system can provide tools for finding them.

Computations and objects

Siren is a pure Object-Oriented language. Users design programs by describing how objects interact with each other. All actions in the system must be carried by sending messages to objects (there are no control flow structures such as if statements or for loops). While this model of programming can be fairly powerful, it's not immediately obvious how one should go about designing these interactions.

In Siren, there are guidelines for effectively designing them. A design process starts by determining the problem, figuring out which components this problem has, and then determining which kind of concept each component should be represented as in the Siren world.

Kinds of concepts can be divided in:

  • A Process is something that, given some data, manipulates it in some way. Note that manipulating data is not the same as mutating data. Whether a process mutates things or not is a separate concern from what said process does. For example, sorting is a process. Writing a file is another process.

  • An Algebra is some kind of entity and a set of processes that manipulate this entity in some way. Data structures are naturally algebras, but any concept might be captured as an algebra. For example, an ATM may be represented as an algebra with operations deposit and withdraw (again, the presence of side-effects in those operations is completely orthogonal to this modelling).

  • An Universe is the domain in which these concepts are presented. So, for example, if you're modelling Trilean logic in Siren, then your universe would be Trilean logic, your algebras would be True, False, and Unknown, and your processes would be And, Or, Invert, and Implies.

In Siren, processes tend to map to messages or blocks, algebras map to objects, and universes map to objects, modules or packages¹. There are cases where that might differ, and, for example, an object might end up representing a process (state machines and actors are good examples of this).

Designing algebras

Once you've figured out the concepts in your universe, you can start splitting them into algebras. First, you need to figure out whether a concept is a single thing (scalar), one thing in many (sum), or many things (product).

Integers are an example of a scalar concept. They're a single entity which makes sense on its own, and can not be broken into further entities² in Siren. Even if an Integer was made up of different objects, those objects are not visible to the user, and the user can not extract those components, or construct an Integer from those components.

Linked lists are an example of a sum concept. They're an entity where each instance of the concept might either be the end of the list, or a node that contains a value, and a pointer to the rest of the list. Booleans are another example of a sum concept, where an instance may be either True or False. While "Linked List" and "Boolean" are sum concepts, each instance of those concepts will have its own classification. Still following the example, end of the list, True and False would be a scalar, whereas a node containing a value and a pointer to the rest of the list would be a product.

Arrays are an example of a product concept. They're an entity where each instance contains many different concepts. Records (e.g.: An entity containing a person's age and their name) are another example of a product.

Making illegal states un-representable

If you're not familiar with the concept of sum types (which is not surprising, given that most programming languages, including Siren, only offer product types), you might be asking yourself why one would need them.

Sum types help with modelling possibilities (Which possible states this entity can be into? Which states can this entity transition to?) naturally, and this helps with correctness because if you capture the possibilities correctly as sum types, it's not possible for your program to enter an illegal state, because illegal states do not exist in your world to begin with.

Consider the previous example of Linked Lists. If one were to design a Linked List without the use of sum types, it would look like this:

let Node = {
  def as-string
    (this empty?) then: {
    } else: {
      this value as-string + ", " + this next as-string

  def empty?

let a = Node { def empty? = True };
let b = Node { def value = 1;; def next = b };
b as-string
(* => 1, (Nil) *)

This object expects two pieces of state: value and empty?. If empty? evaluates to True, then value is not expected to be present. Otherwise, value is expected to be present. In the example usage, the nodes are correctly constructed, so sending as-string to the b object works. But consider the following:

let c = Node { def value = 1;; def next = Node };
c as-string

This fails because Node has empty? evaluating to False, but does not have a value message. Furthermore, the definition of as-string is complicated because it needs to deal with these conflicts of possibilities, and extending this object to support additional states would only complicate matters further. The recommended approach is to, instead, factor these possibilities into distinct objects with the relevant messages in them.

Note: Siren does not support sum types for now, one must encode them by using dynamic dispatch with product types. It's not clear if there's a way of Siren to naturally support sum types without introspecting objects and complecting messages, or if it would be an advantage to do so. This remains an open problem.

let Linked-List = {
  (* Ways of constructing values of this universe *)
  def empty
    this traits Empty

  def of: value
    this empty, value

  (* Shared behaviour in this universe *)
  def traits {
    def Empty = Empty
    def Node = Node

let Empty = Linked-List {
  def as-string

  def , value
    this traits Node {
      def value = value
      def next  = this

let Node = Linked-List {
  def , value
    this traits Node {
      def value = value
      def next  = this

  def as-string
    this value as-string + ", " + this next as-string

let a = Linked-List empty;
let b = a, 1;
b as-string

In the revised example not only is it impossible to create a Linked-List in an illegal state, since you need to start from one of the available constructors in the Linked-List object (assuming Empty and Node are not exposed globally), and all the possible transitions to other states are constrained to be correct by their relative traits, the whole solution is more extensible. One could clone Linked-List and define different behaviours for Empty and Node if necessary.

Interlude: A Review of Siren's Object Model

Siren is a pure object oriented language, which largely means that the entities available to model the world in Siren are objects. In Siren, everything is an object, and all of the operations work on objects. This gives Siren a sufficiently expressive computational and compositional power — the result of composing an object is another object, which can be further composed with existing operations into further objects.

An object is an entity together with the set of operations it supports. In Siren all objects carry all of the operations that they support. The only way of "using" (or rather, interacting with) an object is to send that object a message. The object might then choose whether to perform the operation (if it can understand it) or not.

This is fairly similar to the way people interact with a computer using the keyboard, for example. Pressing one key means sending a message to the computer, which then might get interpreted as a character and displayed on the screen, or might cause a menu choice to be made (in case there is a shortcut defined for that key). The message is how the user specifies which operation they want to carry out, and the operation is how the target decides to interpret that message.

In siren operations are a collection of Siren statements, which are executed in order if the operation is to be carried out. Messages are unique identifiers given to these operations. A message name is a way for someone to refer to these operations. Unlike most existing Object Oriented languages, messages and message names are separate concepts. A message is guaranteed to be globally unique, which means there can never be a collision between two messages in Siren. A name, on the other hand, is a way of resolving these messages. Names in Siren are lexical and defined by the call-site.

For example, in:

let unit = {
  def as-string

  def + that

  def foo: x bar: y

as-string, + and foo:bar: are message names. A unique identifier for those messages is created behind the scenes, and the names are added to the global mapping of Name → Message, so people can send messages to the unit object.

Sending a message to this object might look like this:

unit as-string;
unit + 1;
unit foo: 1 bar: 2

When Siren executes these operations, it figures out which message as-string means for the unit object in the current scope, then sends that message to the unit object. The unit object then takes over the process and figures out which operation that message means, finally it returns "<unit>" as the result of carrying out the operation defined by that message.

A user might override the mapping of names to messages by using the use expression in the language:

let my-unit-as-string = extend unit with {
  def as-string "<my-unit>"

use my-unit-as-string in {
  unit as-string;   (* => "<my-unit>" *)

unit as-string  (* => "<unit>" *)

Translating Algebras into Siren

At this point you should have a set of (Entity × Processes) ∈ Universe that feels consistent. What's left is representing this using Siren's concepts. As a pure object-oriented language, most of these concepts will map to objects. This might sound unnatural, but having a single concept means we have a much greater compositional power.

While we've got many concepts mapping to a single concept, objects can be broken down into several different categories depending on what they're used for. The major categories in Siren are Traits, Examples and Refinements.

  • A Trait is an object that provides a (possibly incomplete) set of features to some entity. Traits are not usable on their own, they must be added to an object that provides the operations it expects.

  • An Example is a fully working object. While Examples can be used on their own, they are not meant to, rather they're provided so people can experiment with them and build on top of them. These are called "prototypes" in OO literature.

  • A Refinement is an example that has been further refined into doing whatever the user needs it to do (e.g.: by providing the object with their data, such that it computes things that are useful to the user). OO literature does not distinguish between "example" and "refinement".

There are a few courses of action when translating algebras to objects, depending on the algebra. But in the common case:

  • Common operations are stored in a trait object.

  • One or more examples of the algebra are provided as starting points to the user, by refining the trait object. Examples are expected to be fully working, even if they don't do anything useful.

  • An entry point object might be exposed to the user which controls how the examples / traits are initially refined. This helps with avoiding the possibility of illegal states.

  • If it doesn't make sense for an object to be refined, and it provides all of the parts it needs to work, then it's not necessary to separate it into a traits object and an example object. Self calls these objects "Oddballs," Siren calls them "Singletons".


  • ¹: So far there are no concepts of packages in Siren, I do not know if we'll ever have them either.

  • ²: This is not entirely true, as you can break integers into bytes, and further into bits. But this underlying representation is opaque in Siren so the user can't see it.