Skip to main content
This document is for maintainers and contributors to Bun, and describes internal implementation details.
The bindings generator scans for *.bind.ts files to find function and class definitions, and generates glue code to interop between JavaScript and native code. There are currently other code generators and systems that achieve similar purposes. The following will all eventually be phased out in favor of this one:
  • “Classes generator”, converting *.classes.ts for custom classes.
  • “JS2Native”, allowing ad-hoc calls from src/js to native code.

Creating JS Functions in Rust

Given a file implementing a simple function, such as add:
src/jsc/bindgen_test.rs
use crate::{JSGlobalObject, JsResult};
use crate::r#gen::bindgen_test as generated;

pub fn add(global: &JSGlobalObject, a: i32, b: i32) -> JsResult<i32> {
    match a.checked_add(b) {
        Some(v) => Ok(v),
        None => {
            // Binding functions can propagate out-of-memory and JS exceptions
            // directly; other failures (like this integer overflow) must be
            // converted into a thrown error. Remember to be descriptive.
            Err(global.throw_pretty(format_args!("Integer overflow while adding")))
        }
    }
}
Then describe the API schema using a .bind.ts file. The binding file goes next to the Rust file.
https://mintcdn.com/bun-1dd33a4e/JUhaF6Mf68z_zHyy/icons/typescript.svg?fit=max&auto=format&n=JUhaF6Mf68z_zHyy&q=85&s=7ac549adaea8d5487d8fbd58cc3ea35bsrc/jsc/bindgen_test.bind.ts
import { fn, t } from "bindgen";

export const add = fn({
  args: {
    global: t.globalObject,
    a: t.i32,
    b: t.i32.default(-1),
  },
  ret: t.i32,
});
This function declaration is equivalent to:
/**
 * Throws if zero arguments are provided.
 * Wraps out of range numbers using modulo.
 */
declare function add(a: number, b: number = -1): number;
The code generator emits a C++ thunk that validates and coerces the JS arguments, then calls the Rust implementation. On the Rust side the generated module is reachable as crate::r#gen::<basename> (for bindgen_test.bind.ts, that’s crate::r#gen::bindgen_test). To construct a JSFunction wrapping the native implementation, use generated::create_add_callback(global):
use crate::r#gen::bindgen_test as generated;

let js_fn: JSValue = generated::create_add_callback(global);
JS files in src/js/ may use $bindgenFn("bindgen_test.bind.ts", "add") to get a handle to the implementation. Exported bindgen functions are snake_cased on the Rust side (requiredAndOptionalArgrequired_and_optional_arg), and the generated callback constructor follows the same convention (create_required_and_optional_arg_callback).

Strings

The type for receiving strings is one of t.DOMString, t.ByteString, and t.USVString. These map directly to their WebIDL counterparts and have slightly different conversion logic. Bindgen passes bun_core::String to native code in all cases. When in doubt, use DOMString. t.UTF8String can be used in place of t.DOMString, but will eagerly convert to UTF-8. The native callback receives a &[u8] slice (WTF-8 data) that is freed after the function returns. TLDRs from the WebIDL spec:
  • ByteString can only contain valid latin1 characters. It is not safe to assume bun_core::String is already in 8-bit format, but it is extremely likely.
  • USVString will not contain invalid surrogate pairs, i.e. text that can be represented correctly in UTF-8.
  • DOMString is the loosest but also the most recommended strategy.

Function Variants

A variants can specify multiple variants (also known as overloads).
import { fn, t } from "bindgen";

export const action = fn({
  variants: [
    {
      args: {
        a: t.i32,
      },
      ret: t.i32,
    },
    {
      args: {
        a: t.DOMString,
      },
      ret: t.DOMString,
    },
  ],
});
Each variant gets a numbered Rust function:
pub fn action_1(a: i32) -> i32 {
    a
}

pub fn action_2(a: bun_core::String) -> bun_core::String {
    a
}

t.dictionary

A dictionary is a definition for a JavaScript object, typically as a function input. For function outputs, it is usually smarter to declare a class type so you can add methods and support destructuring.

Enumerations

To use WebIDL’s enumeration type, use t.stringEnum to create and codegen a new enum type. An example of stringEnum as used in fmt_jsc.bind.ts / bun:internal-for-testing:
export const Formatter = t.stringEnum("highlight-javascript", "highlight-javascript-redacted", "escape-powershell");

export const fmtString = fn({
  implNamespace: "js_bindings",
  args: {
    global: t.globalObject,
    code: t.UTF8String,
    formatter: Formatter,
  },
  ret: t.DOMString,
});
On the Rust side, the enum is mirrored as a #[repr(u8)] enum. Note that bindgen sorts t.stringEnum values alphabetically before emitting the C++ enum class, so discriminants must match the generated header’s order, not the .bind.ts declaration order:
pub mod js_bindings {
    #[repr(u8)]
    #[derive(Copy, Clone, Eq, PartialEq)]
    pub enum Formatter {
        EscapePowershell = 0,
        HighlightJavascript = 1,
        HighlightJavascriptRedacted = 2,
    }

    pub fn fmt_string(
        global: &JSGlobalObject,
        code: &[u8],
        formatter_id: Formatter,
    ) -> JsResult<bun_core::String> {
        // ...
    }
}
WebIDL strongly encourages kebab-case for enumeration values, to be consistent with existing Web APIs.

implNamespace

Setting implNamespace: "foo" on a fn({...}) routes the generated call to crate::<basename>::foo::fn_name instead of crate::<basename>::fn_name. Use this to group related binding implementations under a submodule.

t.oneOf

A oneOf is a union between two or more types. It is represented as a Rust enum with one variant per member type.

Attributes

There is a set of attributes that can be chained onto t.* types. On all types:
  • .required, in dictionary parameters only
  • .optional, in function arguments only
  • .default(T)
When a value is .optional, it is lowered to a Rust Option<T>:
export const requiredAndOptionalArg = fn({
  args: {
    a: t.boolean,
    b: t.usize.optional,
    c: t.i32.enforceRange(0, 100).default(42),
    d: t.u8.optional,
  },
  ret: t.i32,
});
pub fn required_and_optional_arg(a: bool, b: Option<usize>, c: i32, d: Option<u8>) -> i32 {
    // ...
}
Depending on the type, more attributes are available. See the type definitions in auto-complete for details. Note that only one of the above three can be applied, and it must be applied last.

Integer Attributes

Integer types allow customizing overflow behavior with clamp or enforceRange:
import { fn, t } from "bindgen";

export const add = fn({
  args: {
    global: t.globalObject,
    // enforce in i32 range
    a: t.i32.enforceRange(),
    // clamp to u16 range
    b: t.u16,
    // enforce in arbitrary range, with a default if not provided
    c: t.i32.enforceRange(0, 1000).default(5),
    // clamp to arbitrary range, or None
    d: t.u16.clamp(0, 10).optional,
  },
  ret: t.i32,
});
Various Node.js validator functions such as validateInteger, validateNumber, and more are available. Use these when implementing Node.js APIs so the error messages match Node exactly. Unlike enforceRange, which is taken from WebIDL, the validate* functions are much stricter on the input they accept. For example, Node’s numerical validator checks typeof value === 'number', while WebIDL uses ToNumber for lossy conversion.
import { fn, t } from "bindgen";

export const add = fn({
  args: {
    global: t.globalObject,
    // throw if not given a number
    a: t.f64.validateNumber(),
    // valid in i32 range
    b: t.i32.validateInt32(),
    // f64 within safe integer range
    c: t.f64.validateInteger(),
    // f64 in given range
    d: t.f64.validateNumber(-10000, 10000),
  },
  ret: t.i32,
});

Callbacks

TODO

Classes

TODO