Ben's Engineering Blog

Introducing Ezno

Published on Thursday 22nd September 2022

Ezno is an experimental compiler I have been working on and off for a while. In short, it is a JavaScript compiler featuring checking, correctness and performance for building full-stack (rendering on the client and server) websites.

This post is just an overview of some of the features I have been working on which I think are quite cool as well a bit on the project philosophy ;)

it is still work in progress, all the examples work but the tool is still in its infancy

Type synthesis and checking

The core of Ezno is a type checker for JavaScript. Type synthesis is analysing syntax and formulating properties of terms. Comparing the information on terms and how those terms are used, a type checker can prevent errors at runtime.

Ezno's type checker is built from scratch. Getting the features I wanted requires a lot of different functionality and needed several new ideas that as far as I know aren't present in any type-system or existing checkers. The checker is fully compatible with TypeScript type annotations and can work without any type annotations at all.

You can think of it as an extension to TSC, similar ideas but taken further

The next few sections go into some unique features of the checker before going into the real benefit of having all this type information.

Dependent typing

One of the key ideas with Ezno is that it attempts "maximum" knowledge of a source. This knowledge includes:

  • Runtimes exception that could happen because of missing properties
  • Code that will never run
  • Expressions that could be collapsed to reduce work
  • Mutations to data

Because of the dynamism of JavaScript, this requires including references to constants (aka known numbers, bools, strings, etc) in the type-system.

While TypeScript includes constant equality, Ezno has built on top of that doing more such as constant operator evaluation.

const x: 5 = 4 + 2;
const x: 5 = 4 + 2;
Type '6' is not assignable to type '5'

Scaling this up you can now see this working for both identifying an invalid property and detecting dead code:

const obj = {
	key3() { return "test" }
}

obj["key" + (2 + 1)]().nonProp;

const five = 5;
if (five + 5 !== 10) {
	// do stuff
}
const obj = {
	key3() { return "test" }
}

obj["key" + (2 + 1)]().nonProp;

const five = 5;
if (five + 5 !== 10) {
	// do stuff
}
Expression is always false, No property with 'nonProp' on { someProp: 2 }

Objects

As well as identifying terms like 4 and "hello", Ezno also treats objects and functions as constant terms:

if ({} === {}) {
	// dead code, two different objects
}

const a = {};
const b = a;
if (a === b) {
	// always true, reference to same object
}

function func() {}
function func2() { return func; }

if (func === func2()) {
	// always true 👀
}
if ({} === {}) {
	// dead code, two different objects
}

const a = {};
const b = a;
if (a === b) {
	// always true, reference to same object
}

function func() {}
function func2() { return func; }

if (func === func2()) {
	// always true 👀
}
Expression is always true, Expression is always true, Expression is always false

Making all function parameters generic/dependent:

While the examples showcase great static analysis of a sequence of statements. Synthesis can be more difficult when it breaks into functions and knowing what happens across call sites. Ezno can trace the flow of data and actions on it by treating every parameter as what most languages refer to as generic:

function addOne(x: number): number {
	return x + 1;
}

const three = 3;
const four = addOne(3);

assertType<4>(four);
function addOne(x: number): number {
	return x + 1;
}

const three = 3;
const four = addOne(3);

assertType<4>(four);

Here the addOne function is annotated that it returns a number. But instead internally in Ezno the actual result is what it synthesised that it returned. The number in the return type is only used as a constraint on the type returned in body. The synthesized return type that Ezno uses in this case is Add<T, 1>. In total Ezno internally views the function as: addOne: <T extends number>(x: T): Add<T, 1>

Untyped parameters

Parameters that don't have a type annotation have inferred constraints based on usage in the function body. The simplest is a function with no constraints:

function id(a) {
	return a
}

assertType<number>(id(2));
assertType<string>(id("Hello World"));
assertType<"x">(id("x"));
function id(a) {
	return a
}

assertType<number>(id(2));
assertType<string>(id("Hello World"));
assertType<"x">(id("x"));
[assertion passed with: 2, 'Hello World', 'x'

Here when the id function is synthesized it infers the a parameter as being generic and thus the function takes the type <T extends any>(a: T) => T.

Inferred generic restrictions:

In the above usage, the constraint of a is initialized to be any and never require any more properties on it. Now moving on to a little bit of a more complex function:

function runMap(obj, func) {
	return obj.map(func)
}
function runMap(obj, func) {
	return obj.map(func)
}

Here initially obj and func are generics that alias any. However, usage of the obj parameter has inferred for the runMap to be type safe obj must have a map property and that map property must be callable with func:

runMap({}, v => v);
runMap({}, v => v);

This is because the .map synthesis changed the alias. The function is synthesized to <T extends { map: (a: U) => any }, U>(obj: T, func: U) => T["map"](U). (where the return type is the call of obj.map with parameter U).

assertType<Array<string>>(runMap(myArray, v => v));
assertType<4>(runMap(
	{ map(cb) { return cb(2) } }, 
	v => v + 2
));
assertType<Array<string>>(runMap(myArray, v => v));
assertType<4>(runMap(
	{ map(cb) { return cb(2) } }, 
	v => v + 2
));

It is important to note that Ezno isn't the first JavaScript type checker that has inferred generics. Hegel infers generics for functions. However I when tried on the above example with its more complicated two levels of inference, Hegel could not figure it out.

The benefit here with generics is that the function can be very expressive and dynamic, but the function still passes off maximum information to the scope where the call occurred.

Hidden function parameters in JavaScript

The usage of this is a hidden parameter to functions, it is specified by the bounded structure rather than arguments at the call site. Here Ezno treats it as generic but separate from actual parameters:

function getThis() {
	return this
}

assertType<Window>(getThis());
assertType<2>(getThis.call(2));
function getThis() {
	return this
}

assertType<Window>(getThis());
assertType<2>(getThis.call(2));

Variables in parent scopes work similary to this. Rather than being passed through the call site it instead exists in a parent of the current environment as:

let a = {};

function x() {
	a.doThing();
}

x();
a.doThing = () => console.log("Hello world");
x();
let a = {};

function x() {
	a.doThing();
}

x();
a.doThing = () => console.log("Hello world");
x();
Calling function requires 'a' to be { doThing(): any }, found {  }

This works using the same generic system that function parameters use. For variables without a type annotation on the variable, it reuses the inferred generic system.

The example above also shows a sneaky feature of Ezno...

Effects / events

One of the problems of JS is that functions can be impure. Impure means it can apply side effects that are not tracked through the returned type.

Ezno tracks side effects that a function may perform:

const data = { x: 0 };

function getFive(obj) {
    obj.x += 1;
    return 5;
}

assertType<0>(data.x);
assertType<5>(getFive(data));
assertType<1>(data.x);
const data = { x: 0 };

function getFive(obj) {
    obj.x += 1;
    return 5;
}

assertType<0>(data.x);
assertType<5>(getFive(data));
assertType<1>(data.x);

Here this function data-wise returns a number with proof of it being equal to 5. But there are additional "effects" that aren't encoded into that return type. When synthesizing functions Ezno tracks mutations through a system in the context/environment called events. Events in a function then get associated with the function referring to the "effects of the function".

assert

The events sequence is sort of a typed intermediate representation and is additionally used for optimisations involving used assignments and such. This has some similarities to SSA. But it is integrated into the checker and is based on types.

Constant functions

Ezno treats function uniquely with a pointer to a function instead of just a "shape" (the same way constant terms and object references work).

We have seen some calculations on operators being calculated at compile time, but the idea carries over to many internal functions. This means it has a direct binding to the function. From this it can definitively know that this function is Math.sqrt allowing the following to work:

let x: 2 = Math.floor(Math.sqrt(5));
let x: 2 = Math.floor(Math.sqrt(5));

This applies to more functions than just that in Math. The following also applies to many other functions in the standard library. Here we see effects, this generics, and constant functions at play:

const myObj = {
	name: "hello world",
	uppercase() {
		this.name = this.name.toUppercase()
	}
};

assertType<"hello world">(myObj.name);
myObj.uppercase();
assertType<"HELLO WORLD">(x.name);

// Also explicit this calling:
const otherObj = { name: "make me upper" };
myObj.uppercase.call(otherObj);
assertType<"MAKE ME UPPER">(otherObj.name);
const myObj = {
	name: "hello world",
	uppercase() {
		this.name = this.name.toUppercase()
	}
};

assertType<"hello world">(myObj.name);
myObj.uppercase();
assertType<"HELLO WORLD">(x.name);

// Also explicit this calling:
const otherObj = { name: "make me upper" };
myObj.uppercase.call(otherObj);
assertType<"MAKE ME UPPER">(otherObj.name);

JSX

Ezno has support for JSX syntax. A key part of this implementation is that it treats different tags as different tags:

assertType<HTMLParagraphElement>(<p>Paragraph tag</p>);
assertType<HTMLParagraphElement>(<p>Paragraph tag</p>);

It also records more information about the relation of elements to interpolated data and event listeners. (Using the same system that Ezno uses to identify unique objects).

const a = <p onClick={(ev) => {
	assertType<HTMLParagraphElement>(ev.currentTarget);
}>Paragraph tag</p>;
const a = <p onClick={(ev) => {
	assertType<HTMLParagraphElement>(ev.currentTarget);
}>Paragraph tag</p>;

You can think of these as being similar to objects with properties. But in this case, the children binding follow through:

function LazyImage(href: string) {
	return <img href={href} loading="lazy">;
}

const myImage = LazyImage(myInput);
assertType<true>(myImage.getAttribute("href") === myInput);
function LazyImage(href: string) {
	return <img href={href} loading="lazy">;
}

const myImage = LazyImage(myInput);
assertType<true>(myImage.getAttribute("href") === myInput);

Which leads on to the reason for adding these bindings:

The "framework"

One of the biggest uses of JS is with declarative user interface programming.

React is one of the most popular libraries for declarative interfaces. However, its internals are computationally expensive. With Ezno one of the goals was to provide enough static analysis to work with the expressiveness of React but use the information to make it do a lot less work.

"framework" is a temporary name for the plugin written on-top of Ezno, Ezno is not the framework. The "framework" is a plugin using logic from the parser and checker and is entirely detachable.

Eliminating the need for a Virtual Document Object (VDOM)

Firstly a definition of the VDOM (from web definitions):

The virtual DOM is a structure akin to the DOM. It is slimmer and has a subset of the API of the structures defined in the DOM JS spec e.g. HTMLElement. VDOM is a virtual representation of the document, actual DOM references the document (e.g. .click() isn't on VDOM structures).

It is used to add to or update the existing actual DOM / UI.

VDOM implements declarative programming. When state changes it recreates the UI by rerunning the method with the declaration method. This allows programming as a map state of the rather than manually adding the imperative updates to the document at every state change.

The VDOM isn't free:

Before we go into eliminating it we have to deal with the why? As well as the downsides that should occur in the alternative.

Every update requires evaluating the UI

Every time an update to state happens, the runtime needs to rebuild the UI whole tree. This is necessary to find that any of the nodes have changed, but the majority of the new UI hasn't changed and you have to create and store duplicate nodes. The UI can require evaluating expensive calculations that often returns the existing value.

Memoization partially solves this, parts of the tree can be wrapped and when compute them it looks for cached versions of the computation. While this solves the recreation time there is additional memory overhead to hold the cached results and hashing and lookup is not free. It is also developer overhead to realise what is static and independent. Extra explicivity-ness in the syntax (unless a auto memoising compiler was added), just overall more complicated.

Memory, computation and library overhead

  • Because x.innerHTML = newUI() doesn't really work because it has to create new elements on browser and thus can lose state of old elements. Instead it uses the VDOM intermediate representation to work out updates actual to DOM nodes rather than fully replacing them. Holding this intermediate representation requires additional memory as now there is both a browser UI tree in memory and the framework interpretation.
  • To find the updates it uses a process named "diffing". After a new frame has been built it has to be compared to an original frame, which requires walking these trees and comparing nodes.
  • And to do all this handling with the VDOM the library has to be shipped with a reconciler on top of the state handler and the diffing algorithm, all leading to larger JS payloads.

Adding event listeners to existing / server rendered markup is more expensive

Unless using event listener delegation ahead of time, when adding event listeners to existing markup VDOM requires a representation of the whole tree to find event listeners. And even if the event listeners are delegated, a VDOM framework requires creating the existing UI tree to perform updates on an existing server-rendered tree, which requires all loading state that only may be read by the UI.

VDOM without the VDOM:

Again, One of the questions for this project was 'what analysis do you need to do to use JSX and the full range of JS expressions that React works with, without running into the complexities and execution expensiveness of a VDOM'.

Ezno runs on fine-grained reactivity directly on the DOM. Because of the tracking of poly types, useState calls return value is encoded in a special variant of generic effectively "dependent on when the function it is in is called". From this, the left and right values are tracked back to the useState call.

const [value, setValue] = useState("");

//                value : any (via dynamic constraint)
//              / [0]
// useState("") 
//              \ [1]
//                setValue : (any) => any
const [value, setValue] = useState("");

//                value : any (via dynamic constraint)
//              / [0]
// useState("") 
//              \ [1]
//                setValue : (any) => any

Because of this, we have three unique types created from this statement that can be tracked.

return <>
	<input onChange={ev => setState(ev.target.value)}>
	<p>{value}</p>
</>
return <>
	<input onChange={ev => setState(ev.target.value)}>
	<p>{value}</p>
</>

Building a compiler to generate imperative calls is very simple:

  • Find calls where a function calling is on a type like useState()[1]
  • Lookup element points (attributes, children, etc) that are of a type where the base useState is the same useState()[0] (type synthesis stores reverse references on objects)
  • Replace setState call with direct DOM updates that affect the element points

These updates are relative to the elements in the DOM as these relationships are found during the type synthesis. Therefore the above would generate something like:

input.addEventListener("change", ({ target }) => {
	target.nextElementSibling().innerText = target.value;
});
input.addEventListener("change", ({ target }) => {
	target.nextElementSibling().innerText = target.value;
});

Here the type information enables generation of this direct update instead of relying on a VDOM at runtime to find this change.

State objects:

const value = useProxy({ count: 2 });

function updateCount(value) {
	value.count++;
}

return <span>{count}</span>
const value = useProxy({ count: 2 });

function updateCount(value) {
	value.count++;
}

return <span>{count}</span>

This sort of thing works in valtio and is the previous implementation in my old framework Prism. This is possible using Proxies at runtime, but are a no-go really for a number of reasons.

Because of effect tracking in Ezno, it becomes fairly straightforward to find state mutations across functions and so these updates can be found and inlined similar to the above. The uniqueness of functions here enable special handling of .map and .push as they are known to be array method after the type synthesis.

const value = useProxy({ items: [] });

return <ul onClick={() => value.items.push(new Date())}>
	{value.items.map(item => <li>{item}</li>)}
</ul>
const value = useProxy({ items: [] });

return <ul onClick={() => value.items.push(new Date())}>
	{value.items.map(item => <li>{item}</li>)}
</ul>

From this, the logic can work out the push call appends a new element to the ul. Therefore it can skip over needing keys and finding list differences in this scenario and generate a simple .appendChild call.

Universality

Most frameworks enable some sort of server side rendering (often abbreviated to SSR). SSR is the functionality in the framework to generate a HTML string representation of the same UI under a state identically to the result on the browser. It is also used in static site generation where the build tool runs all the requests at build time and saves the results to some sort of static output.

Ezno compiles JSX trees to string concatenation methods. Most compiled frameworks can do this and build alternate functions to what the client deals with.

Here are some of the features and problems at play...

Hydration:

Hydration is a bit of a thrown-around word but it really refers to how to get "state" into some existing object.

One problem is: where does this state come from...?

The double data problem:

One thing with a lot of frameworks is that server-side rendered pages send down JSON blob with the state. The site effectively has to server-side render twice. Once mapped into HTML and secondly as a JSON blob. JSON serializing and deserialization is not free:

A larger document leads to a longer time to generate on the server and a longer time to parse on the client. GZIP / brotli compression reduces the size over network and the size of the cached document as it can mark and reference duplicated strings. However it doesn't completely zero the costs because the keys and other JSON related structure is not in the DOM. If you only want one property of the state it requires deserializing everything.

These problems are also true for other serialization formats not just JSON. So where else can you get state? Well most of the state is rendered in the initial UI as HTML. If the type information knows how state maps to UI the reverse of this information could be used to get state on the client using the initial HTML. This way the JSON blob could be dropped reducing the size of the document!

Lazy state loading mechanism

Aside from size, with a monolithic JSON string object, everything must be deserialized at once.

Because Ezno reads from values in an existing parsed tree the properties can be independently hydrated. Additionally Ezno can do this lazily only when the value is actually read rather at initial load.. Because of the type information, it knows how to convert the string and other representations of data in the DOM on the ones necessary on the client. Therefore generating something like:

const [a, setA] = useState(0);
<button onClick={() => setA(a => a + 1)}>{a}</button>
const [a, setA] = useState(0);
<button onClick={() => setA(a => a + 1)}>{a}</button>
button.addEventListener("click", ({ target }) => {
	target.innerText = parseInt(target.innerHTMl) + 1;
	//                 ^^^^^^^^^^^^^^^^^^^^^^^^^^
	// Here is the reverse expression. Again it knows about types so can invert the implicit cast of a integer to a string
});
button.addEventListener("click", ({ target }) => {
	target.innerText = parseInt(target.innerHTMl) + 1;
	//                 ^^^^^^^^^^^^^^^^^^^^^^^^^^
	// Here is the reverse expression. Again it knows about types so can invert the implicit cast of a integer to a string
});

Reversibility and the fallback for non-injective UI

Firstly only data that is read needs to be hydrated on the client. For state that needs to be read but isn't directly in the HTML, it gets a little bit more complicated. Some expressions like y = x + 1 can be easily reversed, expressions such as Math.sin(x) require a bit more information but with the information about unique functions it isn't impossible.

For data not present in the UI or not reversable that needs to be on the client Ezno falls back to adding this data to attributes of local elements.

This is the least amount of data that needs to be added to the body for it be interactive. Using localized attributes instead of a JSON blob enables lazy and partial state loading.

Downsides of reversibility

Aside from being complex to implement, there are some non-obvious downsides. Firstly lazily pulling data from the HTML can be slower than pulling it from a eagerly hydrated JSON blob, especially if it requires reverse transformations. Also logic for retreival increases the bundle size a little bit.

Page loading, adding event listeners and the double run problem

For upgrading existing HTML, rather than each element registering its own event listener (which can result in hundreds of addEventListener calls) it instead adds a single one at the top level. The output of the framework is still a work in progress but with the bijective property map the position of elements works without the notion of components.

createRoot(document.body).render(() => <>
	<button onClick={() => console.log("clicked")}>Log click</button>
	<button onClick={async () => { await fetch("/do-thing") }}>Make request</button>
</>);
createRoot(document.body).render(() => <>
	<button onClick={() => console.log("clicked")}>Log click</button>
	<button onClick={async () => { await fetch("/do-thing") }}>Make request</button>
</>);

possible compiled output:

document.body.addEventListener("click", async ({ target }) => {
	const id = target.getAttribute("data-click");
	if (id === "0") {
		console.log("clicked")
	} else if (id === "1") {
		await fetch("/do-thing")
	}
});
document.body.addEventListener("click", async ({ target }) => {
	const id = target.getAttribute("data-click");
	if (id === "0") {
		console.log("clicked")
	} else if (id === "1") {
		await fetch("/do-thing")
	}
});

Other ideas:

Some other ideas that type information enables:

Server side rendering out of the JavaScript runtime

One of the problems with JavaScript frontend server rendering implementations is that SSR / "string builders" run on JavaScript which means they are locked into a server runtime like Node or Deno. This is perfectly fine for most things but I think this is generally quite a heavy restriction, which prevents using a lot of cool backend technologies written in other languages like Rust, Python etc. Some tools have embeded a JavaScript runtime into the application (such as rusty_v8). However joining these two systems together can lost type safety bridge and data often has to be copyied into the runtime which seems to void the performance improvements.

With type information it could allow Ezno to generate some format that is tightly integrate with the server language. This is quite simple to do for string elements and Ezno knows the shape of data and the pointer to every function it could easily transpile some of this stuff. However it is still up in the air how this would work if the server has to do something more complicated mutating data and I don't have a clear idea of what the format would be that could be used across languages would look like?

Linting

With types you can restrict the code written. For example Math.sin is typed as being (x: number) => number, however JavaScript is more flexible and Math.sin([4]) is still a valid call even though [4] is not a subtype of number.

With type information you can (optionally) check and deny some behavior based on the values of things. This could be used to enforce semantic HTML. e.g imgs require a alt tag property, functions registered for the click event cannot make navigation calls unless they are a anchor tag...

There is quite a lot a more you can do with a type system other that checking you have spelt JSON.stringify correctly and using the information to optimising things to run slightly faster.

Auto progressive enhancement

As Ezno knows at compile time what functions have been bound to certain event listeners plus the internal effects it knows to extract runtime state mutations to the server to add a backup to client side interactions if JavaScript fails to load.

Plugins and extensibility:

Ezno is written in Rust and has several places/hooks for adding additional functionality:

Exterior type safety:

Because Ezno treats functions as unique it allows for special handling of functions via Ezno's plugin system.

For example, the fetch function could be overwridden and for known strings could return a more precise type based of knowledge of what a endpoint returns.

Build tool front-ends:

Aside from the CLI, there is a language server plugin (LSP) so you can use it in a editor. For those who don’t like the command line build step, there is a half-working Nodejs runner, this means you can integrate the build step into the runner compressing running the source into a single command. For the web there is a WASM service worker runner that shouldn't need to touch the filesystem during development.

Complexity & wrapping up the features

Enough with the features here is a roundup and some philosophy of the project

Types and TypeScript

Types are an integral part of JS with things like instanceof and typeof so I don't think some of these ideas are too foreign with the language. Type annotations in source code are treated as restrictions under Ezno and so are not strictly necessary. This differs from TypeScript which treats things like the return annotation as the source of truth, allowing this to be validly checked:

// Passes under tsc
function y(a: any): string {
	return a
}

y(2).slice(2)
// Passes under tsc
function y(a: any): string {
	return a
}

y(2).slice(2)

TypeScript has a bit of a funky implementation around any allowing a to be cast as a string in the above example. Implementing any this way makes TypeScript easier to adopt and allows things to compile in weird environments. However for Ezno to do its optimisations this magic any type that has the property of all types without narrowing doesn’t quite work.

From the previous section on events and some notions around generics, understandably, the information needed is more complex (and verbose) than simple annotations allow, thus treating Ezno types as the source of truth rather than annotations in the source code. However, some of these annotations are possible. To describe the "root environment" of the JS standard library, Ezno uses modified TypeScript definition files peppered with decorators that register information associated with functions such as function bindings and effects they can run.

Don’t take this as a knock on TypeScript, TypeScript is great. Ezno started off as pet project to re-implementing the features of TSC in Rust. But as things started to go well after adding more and more TSC features, it was apparent that bigger such as hydration and the reactive system goals weren’t going to be possible directly following its path. It isn’t possible to detect what can happen to state when called through a term that represents any. TypeScript holes aren’t particularly critical if you just want code completions and some level of type safety. They allow the more complex parts to not be blocked by compiler errors. But for doing optimizations a single unknown result can make it impossible.

Currently Ezno isn’t a feature-complete type checker. There are still a lot of things to still work on. For example Ezno "proof via predicates" is in the works to add to the existing "proof by definition" type system. So far the first section goes over the analysis on completely static code. The next steps are to apply these ideas to more dynamic structures. There are also some ideas not mentioned here to prevent this post from being too long 👀.

Frameworks:

The output of the framework is still work in progress. There is still a few things to add to the type checker that are necessary to get "the framework" plugin to work in edge cases. Thus no benchmarks or definite output in this post. Most benchmarks show hand written JavaScript code to be the most efficient so the idea is to take the syntax and information about it to squeeze it into the closest of the hand written forms. The performance gains from all this optimisations are probably un-noticeable. This is really an exercise in attempting to get to the least amount of code to run to make a page interactive.

Some frameworks abstract through libraries and structure of functions. Some compiler-based frameworks abstract through syntax. Ezno is more unique it is based on data and the semantics of the program above what is visible from the syntax.

Ezno today

I am still unsure on the name. I think it is fun short and quirky and most importantly, not taken on package managers. It has a little bit of a hidden meaning ("easy? no" referring to doing static analysis on JavaScript).

I want to keep things moving but slowly. On initial viewing incomplete projects seem bad but I think that there is a good as they have space to add additional features and improve. Some projects I see which seem in closer to their goal have problems with improving. I think some tools are built too quickly and I don’t want Ezno to fall into that category. It is a bit earlier than I want to speak about the project but I am keen to see feedback about what issues this fixes as well as other ideas to persue.

No demo binary out yet, need to finish of some things on more advanced events to get some of the cooler demos to work. Hoping for something demonstratable before the end of this year 🤞