Skip to content

Latest commit

 

History

History
531 lines (407 loc) · 17.3 KB

README.md

File metadata and controls

531 lines (407 loc) · 17.3 KB
Raw

ValueScript

A dialect of TypeScript with value semantics.

Playground

Try ValueScript instantly using your web browser.

About

ValueScript uses TypeScript syntax, but it compiles to a bytecode that runs in a different virtual machine.

The syntax is identical, not just similar. We use SWC's TypeScript parser. This means you can use your IDE's TypeScript functionality when writing ValueScript.

This program shows the core difference between ValueScript and TypeScript:

export default function main() {
  const leftBowl = ["apple", "mango"];

  let rightBowl = leftBowl;
  rightBowl.push("peach");

  return leftBowl.includes("peach");
  // TypeScript:  true
  // ValueScript: false
}

In TypeScript, "peach" is in the left bowl because TypeScript interprets rightBowl = leftBowl to mean that there is one bowl and both variables point to the same bowl. That one bowl is changed by .push("peach").

In ValueScript, objects never change this way, only variables change. Pushing onto rightBowl is interpreted as a change to the rightBowl variable itself, not the data it points to.

You can see this in the playground, or run it locally:

git clone [email protected]:voltrevo/ValueScript
cd ValueScript
cargo build -p vstc
export PATH="$PATH:$(pwd)/target/debug"
vstc run inputs/passing/readme-demo.ts

One way to understand this is to imagine that things like = and passing parameters are implemented by deep copying. We don't implement it that way, but it would work the same if we did (just a lot slower).

Instead, we implement rightBowl = leftBowl by sharing the memory, but there's also a reference count attached to that memory. When rightBowl is mutated, the VM can see that it's using shared memory, and does some bookkeeping to represent rightBowl's updated value without mutating the shared memory.

By the same token, if leftBowl had gone out of scope or was optimized away, the VM would see that rightBowl has the only reference to that memory, and would mutate it directly.

No Side Effects

ValueScript has no side effects, with two exceptions:

  1. You can (in future) choose to introduce side effects via foreign functions
  2. Bugs (please report them)

ValueScript does this by behaving differently to TypeScript in three key ways:

  1. Value semantics (see About)
  2. Captured variables can't be mutated (mutation is otherwise encouraged)
  3. When this changes inside obj.method(), the updated this value is used to mutate obj when the method returns

(2) and (3) are described in more detail in this article.

Intended Usage

ValueScript has its roots in a programming school of thought that discourages the use of mutation.

While this idea is indeed useful, the result of this approach can only push the mutation to the edges of the program. At the edge, you need to interact with things that are inherently mutable, like users. The code you write becomes a subsystem of some larger framework that dictates the interface with the outside world.

This is why we expect that ValueScript will be most useful as a tool within a TypeScript project, rather than an alternative to it. This way you can benefit from TypeScript's rich ecosystem to interact with users and external systems, and also have a clearly separated immutable subsystem to define the core of your application.

Because ValueScript shares the same syntax as TypeScript, you'll be able to inline ValueScript code into TypeScript like this:

const points = inlineValueScript(() => {
  const x = [3, 5];
  const y = x;
  y[0]--;

  return { x, y };
});

// ValueScript doesn't have console.log, but TypeScript does.
console.log(points); // { x: [3, 5], y: [2, 5] }

Additionally, ValueScript has benefits that make it a suitable target for secondary storage. Rather than writing to a file, your ValueScript code can read and update persistent objects the same way it interacts with regular in-memory objects.

Benefits

Eliminate mutation bugs

Mutating things across your program is frequently intended, but it's also frequently unintended, causing bugs.

This is why you are usually encouraged not to mutate function arguments, among other things. Sometimes you'll see workaround like const a = [...b];. In ValueScript, just write it the natural way.

`const` means what you think it does

Ever felt weird about using const in situations like this?

const values = [];

values.push(123);

return items;

Us too. The reason is that, in a mutable world, it's the array that values points to that is mutating. Pushing to that array doesn't change values - it still points to the same array, right?

In ValueScript, it's not the same array, because arrays don't change. Instead, it is indeed the variable that changes, and therefore, if you mark it as const, attempting to do so is a compile-time error.

Testable code

Testing code is all about being able to draw a boundary around something that can be given inputs so that you can check its outputs against your expectations.

Being able to draw these boundaries is usually challenging in real-world systems, because by default everything wants to connect to something tangible to serve its purpose as directly as possible. Most things that matter to you become untested because of their coupling to externalities that are too difficult to meaningfully replicate in a test case. Testing degrades into an inauthentic add-on that focuses on trivialities.

By using ValueScript, you can maintain a clear separation between a domain that should be easy to test - the core of what your application does, and a domain that is difficult to test - how your application talks to the world.

A ValueScript program is always a function that, when called with the same inputs, produces the same outputs.

No garbage collection

In ValueScript, it's impossible to create data that circularly references itself. This isn't because something is keeping watch and producing an error if you do it accidentally. Rather, it's just an inherent consequence of how ValueScript works:

let x = {};
x.x = x; // { x: {} }

// (In TypeScript: { x: { x: { x: { x: { ... } } }} })

Circular references are the whole reason why garbage collectors are needed (assuming you want to reuse memory and don't want to figure out when it's safe to do so). Without them, ValueScript is able to simply keep a count of how many references each object has, and when that count drops to zero, it cleans up the memory immediately.

Persistence

In a traditional mutable program, the important entities in that program often can't be stored authentically without also capturing the state of the entire program that contains them. Even when that isn't true, the entity needs to be translated into a form that can be stored in a process we know and love called serialization.

ValueScript is different. Everything can be persisted as its direct contents and a recursive inclusion of its dependencies. This includes functions and class instances (and the methods on those class instances). In ValueScript, everything is plain data.

In fact, because ValueScript doesn't require garbage collection, it's also possible to build up large structures that wouldn't fit into memory. In garbage collected languages, the garbage collector needs to be able to fully traverse all the data (as a last resort) to find cycles to clean up, so growing beyond memory limits isn't very practical. ValueScript doesn't have this limitation.

Make use of TypeScript's type checking

ValueScript is similar enough to TypeScript that the type checker correctly identifies type errors in ValueScript.

In fact, when the differences matter, the type checker often actually favors ValueScript, not TypeScript.

E.g.

let a: { value?: string | number } = {};
a.value = "str";

let b = a;
b.value = 37;

type T = typeof a.value;
//              ~~~~~~~ TypeScript: 37
//              ~~~~~~~ ValueScript: "str"

// The type checker assigns `string` to `T`.
Concurrency

tl;dr:

  • (This is not implemented yet)
  • ValueScript is multi-threaded
  • Calling an async function creates a new thread
  • Because ValueScript functions are pure (async or not), the concurrent evaluation is guaranteed to be the same as sequential evaluation (ie no race conditions)
  • You can write value = promise.wait() in a sync function, because this doesn't block other threads from running

By using value semantics, ValueScript ensures that a function, called with the same arguments, always returns the same value (except for any side effects you choose to introduce with foreign functions). This includes instance methods by considering the instance data to be one of the arguments.

This means that once a function has its arguments, its result is fully determined. It would be safe to evaluate the function concurrently because its output cannot be affected by other code:

const f = (z: number): number => {
  const x = widget.calculate(37);
  const y = expensiveCalculation(z, z);

  return x + y;
};

Above, widget is captured by f. ValueScript requires that captured variables are const, which means that widget cannot change, and therefore widget.calculate(37) cannot change. This means that the value of f(z) is independent of any other work that happens in our program.

Therefore, we could safely evaluate f(z) concurrently. In future, some calculations might automatically be upgraded to concurrent execution, but knowing when it is worthwhile to create a separate thread is a complex and inexact science.

Instead, in the foreseeable future, ValueScript will allow concurrent evaluation of async functions. Even if f isn't already async, you could evaluate it concurrently like this:

const fPromise = (async () => f(z))();

Alternatively, something like vs.thread could make this more clear:

const fPromise = vs.thread(() => f(z));

Of course, functions like f could be made async to begin with, to signal the intent that they are expensive calculations that justify a thread:

const f = async (z: number): Promise<number> => {
  const x = widget.calculate(37);
  const y = expensiveCalculation(z, z);

  return x + y;
});

Now f just returns a promise:

const fPromise = f(z);

Later, when you need the value inside fPromise, you can use await as normal:

const fValue = await fPromise;

However, this requires you to be inside an async function.

In JavaScript, it would be a big no-no to allow a method that synchronously extracts the value of a promise by blocking evaluation until it became available. This is because JavaScript is single-threaded, and there's usually other work the runtime could be doing.

In ValueScript, the other work happens in other threads, so there's no reason to prohibit it. ValueScript allows this via promise.wait():

const fValue = f(z).wait();

Suppose instead that f, widget.calculate, and expensiveCalculation are all sync functions. Suppose that f is part of an important API - it has users and you can't require them to make changes. Allowing .wait means those users can still benefit from this multi-threaded version of f:

const f = (z: number): number => {
  const [x, y] = Promise.all([
    vs.thread(() => widget.calculate(37)),
    vs.thread(() => expensiveCalculation(z, z)),
  ]).wait();

  return x + y;
};

You could also simplify code like f with a utility like parallel:

// Simple version that unnecessarily widens T when the jobs return different
// types
function parallel<T>(...jobs: (() => T)[]): T[] {
  return Promise.all(jobs.map(vs.thread)).wait();
}

const f = (z: number): number => {
  const [x, y] = parallel(
    () => widget.calculate(37),
    () => expensiveCalculation(z, z),
  );

  return x + y;
};
Static Analysis & Optimization

ValueScript dramatically expands the cases where program behavior can be determined statically. In traditional languages, inferences about data in variables are quickly lost because it is impossible to know whether some other code might modify that data.

A relatively simple application of this is tree-shaking. ValueScript analyzers will be able to determine much more accurately what code is actually used, and only include that code for distribution. During development you can also get a lot more feedback like 'this statement has no effect'.

Another important use-case here is testing. In the future, ValueScript will include vs.staticTest(name, fn) which accepts a function taking no arguments, which can therefore be computed statically. The compiler will emit an error if the test fails.

Status

ValueScript is in early development. There may be some descriptions of ValueScript elsewhere here that represent how ValueScript is intended to work, not the subset of ValueScript that has actually been implemented.

Implemented
  • console.log
  • Classes
  • Closures
  • Loops
  • Recursion
  • Destructuring
  • Exceptions
    • Variables changed during try block are reverted on catch
  • Enforcing const
  • Temporal dead zones
  • Local imports
    • Including the many various import and export patterns
  • Tree shaking
  • Copy-on-write optimizations
  • utf8 strings (not JS's utf16 strings)
    • "🫣".length -> 4
    • (JS: -> 2)
    • [0, 1, 2, 3, 4].map(i => "🫣"[i]) -> ["🫣", "", "", "", undefined]
    • (JS: -> ["\ud83e", "\udee3", undefined, undefined, undefined])
  • Math
  • Array standard methods (.sort, .map, .filter, etc.)
  • Most string standard methods (.includes, .slice, .split, etc.)
  • BigInt
  • Iterators
  • Spread operator on iterables
  • Generators
  • Structural comparison
    • {} === {} -> true
    • new Point(1, 2) === new Point(1, 2)
    • (() => {}) === (() => {})
    • JS: -> false
    • This is a value semantics thing - objects don't have identity
  • TypeScript enums
  • TypeScript parameter properties
  • Capturing this in arrow functions
  • Many unusual JS things:
    • [] + [] -> ""
    • [10, 1, 3].sort() -> [1, 10, 3]
    • "b" + "a" + +"a" + "a" -> "baNaNa"
    • (With few exceptions like utf8, the goal is to just do things the JS way, to maximize familiarity for people coming from JS. We're open to revising this strategy, subject to community feedback.)
Not yet implemented

Ecosystem

  • Foreign functions
  • Standardized foreign function packages for web/node/deno-like APIs
    • Sadly, the only access to the host environment is currently console.log
    • We consider this extremely important, but want the language itself to be more robust before embarking on this enormous category of work
    • (Some small & strategic host access will probably be implemented earlier)
  • Tools for embedding ValueScript in other languages
    • (The playground kinda does this, but the solution is purpose-built for the playground, and not intended to be used in other projects (you're welcome to try it of course, but better solutions are planned))
    • Webpack integration
      • import immutableStuff from "ValueScript:./path/to/immutableStuff";
    • Building JavaScript bundles containing ValueScript via embedded WebAssembly (or importing the required WebAssembly)
    • Transpiling ValueScript into JavaScript
      • E.g. a.b.c++ -> a = { ...a, b: { ...a.b, c: a.b.c + 1 } }
    • inlineValueScript(() => { /* ValueScript */ })
      • Uses .toString() to get the source code and compiles and runs it in WebAssembly
    • C libraries, and bindings for python etc
  • Dynamic imports
  • Importing modules from npm
    • (Even when this is implemented, many modules won't work due to their intention to run in a JS environment though. At least at first.)

Core

  • Object spreading
  • Rest params
  • Async functions
  • TypeScript namespaces
  • import.meta
  • Unusual JS things like passing unintended types to standard functions
  • A workaround for JavaScript's utf16 strings
    • jsˋ🫣ˋ.length -> 2
    • [0, 1, 2].map(i => jsˋ🫣ˋ[i]) -> [jsˋ\ud83eˋ, jsˋ\udee3ˋ, undefined]
    • (To be fair to js, note that iteration uses code points: [...jsˋ🫣🚀ˋ] -> [jsˋ🫣ˋ, jsˋ🚀ˋ])
  • JSX
  • Regex
  • Date
  • Stack traces
Not planned
  • Prototype pollution
  • Mutating imported variables
  • Reference semantics
  • Mutating captured variables
  • "Everything is an object"
    • Properties cannot be set on non-objects like arrays and functions
    • new Number() throws a TypeError instead of creating a non-primitive number-like thing
  • The with keyword
  • utf16-based operations on native strings
    • (But see jsˋˋ workaround in not-yet section)
  • Math.random (except as an opt-in foreign function)
  • Date.now (except as an opt-in foreign function)

Contributing

We'd be thrilled to have your help! Please see CONTRIBUTING.md.