inline-functions

inline-functions

This library allows you to inline functions in your JS/TS code. This is useful for performance reasons, as it allows you to avoid function calls and instead inline the function body directly.

It comes as a collection of plugins for different build systems. It is designed to be used with vite, rollup, and esbuild.

Inspired by the clever trick used in robust-predicates by @mourner.

Installation

npm install inline-functions [vite] [rollup] [esbuild]

How it works

This plugin works by replacing function calls with the function body. This is done by using a macro system. The macro system is a simple object that maps function names to their bodies. For example:

import { Vec2, fn2 } from "./utils";

// this is the "real" implementation of the function
// it guarantees that the code will work even if the inlining step is skipped
const add = (a: Vec2, b: Vec2): Vec2 => {
  a[0] += b[0];
  a[1] += b[1];
};

const foo: Vec2 = [4, 5];
const bar: Vec2 = [5, fn2(6, 12)];

const ZERO = (): Vec2 => [0, 0];

const baz: Vec2 = [33, 22];
// note that you can use the function in the macro,
// if your replacement function can handle it
add(baz, ZERO());

// so that dead code removal doesn't remove the function altogether
console.log(foo, baz);

// this will work, but will be removed as dead code
add(ZERO(), bar);

Then the macros to inline the add() function calls, would look like this:

// counter is used to generate unique variable names
let counter = 0;
const macros = {
  add: (a, b) => {
    const _a = `_a${counter++}`;
    const _b = `_b${counter++}`;
    // this will reference the arguments passed to the function
    return `
  const ${_a} = ${a};
  const ${_b}  = ${b};
  ${_a}[0] = ${_a}[0] + ${_b}[0];
  ${_a}[1] = ${_a}[1] + ${_b}[1];
  `;
  },
};

And the result of this replacement is:

var foo = [4, 5];
var ZERO = function () {
  return [0, 0];
};
var baz = [33, 22];
// note that you can use the function in the macro,
// if your replacement function can handle it
const _a0 = baz;
const _b1 = ZERO();
_a0[0] = _a0[0] + _b1[0];
_a0[1] = _a0[1] + _b1[1];
// so that dead code removal doesn't remove the function altogether
console.log(foo, baz);

Usage

Vite

npm install inline-functions vite

Then, in your vite.config.ts:

import { defineConfig } from "vite";
import inlineFunction from "inline-functions/vite";
import { resolve } from "path";

import { macros } from "./macros";

export default defineConfig({
  ..., // your config
  plugins: [inlineFunction({ macros })],
});

Esbuild

npm install inline-functions esbuild

Then, in your esbuild.config.js:

import inlineFunction from "inline-functions/esbuild";
import { macros } from "./macros";
esbuild
  .build({
    entryPoints: ...,
    bundle: true,
    outfile: ...,
    plugins: [inlineFunction({ macros })],
  })
  .catch(() => process.exit(1));

Rollup

npm install inline-functions rollup

Then, in your rollup.config.js:

import inlineFunction from "inline-function/rollup";
import { macros } from "./macros";

export default {
  input: ...,
  output: {
    ...
  },
  plugins: [inlineFunction({ macros })],
};

Verbose mode

For debugging, you can pass { macros: ..., verbose: true } to the plugin in options, and it will print out all the replacements into stdout.

Limitations

Unpredictable amount, because there's no code analysis, it's just a string replacement. The main limitation for now is that the inlined functions cannot return, because it's not obvious how it will fit into the context. You can use the expressions as arguments, but be careful with that. The replacements really have to be simple in order to work efficiently.

Examples

See test for examples and benchmarks.

Some benchmarks

Running "Vite" suite...
Progress: 100%

  Inlined:
    1 464 435 ops/s, ±4.63%   | fastest

  Not inlined:
    1 187 396 ops/s, ±5.75%   | slowest, 18.92% slower

Running "Rollup" suite...

  Inlined:
    1 560 921 ops/s, ±1.27%   | fastest

  Not inlined:
    1 302 231 ops/s, ±0.11%   | slowest, 16.57% slower

Running "esbuild" suite...

  Inlined:
    1 564 889 ops/s, ±0.80%   | fastest

  Not inlined:
    1 281 501 ops/s, ±1.02%   | slowest, 18.11% slower