The Lobster Type System

This document is about how to make most effective use of the type system..

The nature of the type system

Lobster was originally dynamically typed. Not only is this easier to implement, it bring tremendous expressive power "for free", particularly in the way it allows you to write generic code which often requires arcane typing annotations in static languages. The downsides are obvious too: type errors come late or stay hidden, and execution speed suffers.

The type system in Lobster attempts to achieve the best of both worlds, with a type system that retains as much as possible of the dynamic style of writing (little to no type annotations needed, while able to type check a much broader range of code than other languages). It achieves this by using 3 techniques that build on each other:

  1. Type Inference. A well known technique by now, as many language employ it one way or another, and it is an obvious way to reduce type annotations. By itself it is rather restrictive however, as it is limited in the kind of generic code it can correctly type.

  2. Specialization. Instead of struggling to infer or define generic types for everything, we specialize functions to the exact parameter types they are called with. This allows a much wider range of possible code inside generic functions, since they are only ever type-checked with specific types. This of course requires there to be multiple copies of functions, but in practice the code expansion is minimal. Specialized code can be optimized and inlined more aggressively, resulting in significant speed advantages besides the expressivity gains.

  3. Flow Based: Type checking happens in order of function calls, i.e. it is much like evaluating the code, but with types instead of values. This means that when type checking a particular function, it not just knows the exact types of parameters (thanks to specialization), but also the types of free variables, and it knows its callers (a compile time call stack). This means that when a type error happens, you can see in what context the wrong type got created and passed down. Think of it as an earlier dynamic type error. In addition, it takes into account the state of conditional statements, allowing you to avoid unnecessary type errors.

As a result, you are less constrained in the kind kind of code you can write that still type checks correctly, while using less annotations. But with great power comes great responsibility: the less annotations you write, the more a compile time error indeed looks like a run-time error: they are almost as hard to track down. It is certainly easy to start with no annotations, and then gradually add them in places to force type errors to come earlier and be easier to understand. Find your own trade-off.

Writing generic code

In the simplest cases, there is no difference between writing dynamically typed code and typed code. For example a simple function like:

def add(x, y): x + y

is how you create a generic functions in either case. You create a generic parameter by.. not specifying a type!

Now this is already a non-trivial example that will trip up many languages with type inference, since you cannot have the arguments be generic unless the + operation is defined on them, requiring them to have a way to specify the set of types that have this operation specified (see e.g. type classes in Haskell), or simply not allowing it at all (see e.g. generics in Java).

Lobster does not have this problem, since it simply type checks a copy of add for each combination of types it is called with. This then deals correctly with allowing to write add("number: ", 4), taking care of converting 4 to a string and calling string concatenation on the result, without making simpler cases like add(1, 2) slower because of type inspection. In fact, the opportunities for statically optimizing these operations are now greatly increased. This is very similar to C++ templates, but without the clunky syntax.

It is typically "good style" in Lobster to use untyped arguments when a) multiple types are possible/useful, and b) there is no relation between the types of the arguments. Sometimes it is useful to use explicit generic annotations, for example:

def add<T>(x:T, y:T): x + y

Would enforce that both parameters are of the same type, unlike the example above. The type T is automatically deduced from the first argument that matches it (in this case, the first), and then any further uses are required to be that same type, so for example add(1.0, 1) would make T a float, and thus coerce the second argument to float before the function call. In the untyped case above, it would specialize add for (float, int) arguments, and the coercion to float (required by +) would now happen inside the function.

You can even specify the generic type explicitly, for example add<float>(1, 1) would actually coerce both arguments ahead of time. If the function has multiple generic arguments, you don't need to specify all of them, if the remaining ones can be inferred. Caveat: this currently only works where the < directly follows a known function name (to distinguish it from uses of the < operator).

And all of this also works on untyped function declarations equally, since they internally get converted to explicit generics by simply adding generic types A, B, C,.. to them for their untypes arguments.

You can use generic types from surrounding functions, and from any class this function is a "method" of (either declared inside that class, or with the class type as first argument).

The trouble with nil

So far, the type system tries to preserve as much freedom of dynamic typing as possible. One area where it is however more strict than most existing languages is how it deals with nil.

nil (or null in other languages) is usually a default value of all reference types, which is very convenient in expressing uninitialized or optional values. But this convenience comes at a high cost: if you forget to check for nil before you access a value, your program may terminate with a run-time error. Very often we know a value cannot be nil, but we cannot guarantee this, or are lazy to do so. Result: random errors popping up in random places. Not good.

Lobster takes the obvious approach: nil is a separate type that is not a valid value of a reference type. You may still create "nilable" types, but these cannot be accessed (e.g. dereferenced) unless they are first converted to their underlying type:

var a = nil      // a is a nilable of unknown type
if ..:
    a = "foo" // a is a nilable string
a += "bar"    // error: + not defined for nil
if a:         // guaranteed not be nil inside block
    a += "!"  // ok: a is of type string here

As you can see, we can use conditionals to guard code that wants to access a nilable. This seems like a hassle maybe, but think of the benefit: you will never see a nil related run-time error ever again.

And now that nil is a separate type, we can improve our code by not using nil unless absolutely necessary. Many uses of nil in legacy code are lazy, and can be avoided:

The flow based analysis generally understand logical expressions, so if you write if not a: instead, you'll be able to access a in the else block, but not in the then block. Similarly, if you write if a and f(a): then a is available as non-nil both as argument to f and in the following block.

? is generally the way to indicate a nilable type, e.g. string? is a nilable string.

One more thing to notice about nil is that it only knows what exact nilable type it represents once it comes "in contact with" its reference type. That means that if you tried to access a right after it was defined, it would have still have type nil rather than string?. Similarly, if you tried to assign a vector to a after the first assignment, it would complain that only a string? is valid, i.e. a string or nil value.

To bypass that process and force nil to represent a particular nilable type, you can attach a type, like nil:string, which creates a value of type string?. Alternatively you could write nil and a, which makes nil have the type of whatever a has, which may be a generic variable. Similarly, a or nil has the same nilable type, but now the value is actually a rather than nil.

More flow based types

As we saw with nil, a variable may have a different type in certain parts of the code, thanks to the information that a conditional provides.

Another example of this is the is operator:

if a is xyz: a.x

We can access fields of xyz regardless of what type a had outside this scope. is is only true if the types are exactly equal, not subtypes.

Assignment can work similar to conditionals:

var a = nil
if ..: a = float3_0
a.x        // error
a = float3_0  // a is guaranteed non-nil below
a.x        // ok

Note how there are two different cases here: with an assignment inside a conditional, the type checker has to be conservative, and can't guarantee the value is non-nil. But with an un-conditional assignment, we can change the type of a for the rest of the block.

An additional way is to use assert. Use this when you know for sure a value must be non-nil or have a particular type, but you can't structure your code such that this is already obvious to the type system.

Compile-time if-then's.

The flow-based type-checking above gives branches more specific types. But we can do even better: if the condition is statically known, we can avoid type-checking altogether:

def compile_time_if(x):
    if x is int or x is float:
        1 / x
    else:
        x
assert(compile_time_if(1) is int)
assert(compile_time_if("") is string)

The division would be a type error if x is a string, but it compiles anyway since this if is compile-time constant, and the type checker ignores this branch.

If you're used to dynamic languages, you might think: what's the fuss? This would be a compile error in almost all statically typed languages, since they don't cull code before the optimizer. For example with C++ templates can't handle this situation, and that's already more powerful than most forms of generics.

Also note the return type is just the else branch in the string case, it doesn't need to make a type union between the branches anymore.

This is powerful, because it allows you to write generic functions that do subtly different things depending on the input, with no overhead from the conditional, and the ability to operate differently on each case.

You can do something similar with overloading / dynamic dispatch, but that may require more code or may be less efficient.

Generic classes and structs

Besides functions, classes can also be generic, but here we only use explicit generic parameters to do so:

class foo<T>:
    a:string
    b:T

You can now create named specializations using this syntax:

class foo_f = foo<float>

You can now construct a foo with floats, either explicitly with foo_f { "hi", 2.0 }, or you can use the generic type: foo { "hi", 2.0 } and let the compiler pick from all available named specializations. If you wish, you can also construct a specialization directly without naming it: foo<float> { "hi", 2.0 }

Functions written to accept foo arguments will of course work on any of the specializations.

You can even specialize and subclass at the same time:

class bar : foo<float>
    c:int

class bar<T, U> : foo<T>
    c:U

The first specializes and subclasses at the same time, creating a non-generic type. the second version subclasses foo into another generic type, passing on one generic parameter and introducing another.

Numeric Structs (math vector types)

Structs whose fields are either all int or all float have a special status in the type system, as they are accepted as arguments to vector versions of the built-in operators (+ - * / etc) as well as many vector operations part of the built-in functions (e.g. normalize).

You can see the default types for these (vec2, vec3, and vec4) defined in modules/stdtype.lobster. Other types also work with these operations, though (e.g. color).

Confusingly, in math, such N dimensional things are commonly referred to as "vectors", whereas Lobster calls its resizable arrays "vectors". So in Lobster these things are "numeric structs", but on occasion may still also be called "vectors", sorry for the confusion.