Optimize encode and decode with TypedArrays

[jridgewell]

For decode, the optimization is to convert charToInteger to a Uint8Array. Indexed lookups on a TypedArray are faster than property lookups on an object. With that, we decode 1, 4, or 5 ints at a time, instead of looping through multiple times for each char. This leads to a ~20% speedup.

For encode, we switch to a preallocated Uint8Array to store the VLQ data. When needed, we'll grow it, and push the next ASCII byte onto it. At the end, we use a TextDecoder to quickly convert to a string. This leads to a ~2.5x speedup.

---

Running this on my M1 MBP (10 core, 32gb):

Running node v14.18.3 (npm v6.14.15)
decode (latest) x 39,390 ops/sec ±0.20% (94 runs sampled)
decode (proposal) x 47,696 ops/sec ±0.12% (97 runs sampled)
Fastest is decode (proposal)

encode (latest) x 19,949 ops/sec ±0.09% (99 runs sampled)
encode (proposal) x 48,402 ops/sec ±0.93% (93 runs sampled)
Fastest is encode (proposal)


Running node v16.13.2 (npm v8.1.2)
decode (latest) x 39,033 ops/sec ±0.42% (97 runs sampled)
decode (proposal) x 47,220 ops/sec ±0.11% (98 runs sampled)
Fastest is decode (proposal)

encode (latest) x 19,109 ops/sec ±0.13% (100 runs sampled)
encode (proposal) x 49,603 ops/sec ±0.68% (99 runs sampled)
Fastest is encode (proposal)


Running node v17.3.1 (npm v8.3.0)
decode (latest) x 38,961 ops/sec ±0.23% (99 runs sampled)
decode (proposal) x 47,153 ops/sec ±0.13% (98 runs sampled)
Fastest is decode (proposal)

encode (latest) x 19,034 ops/sec ±0.16% (101 runs sampled)
encode (proposal) x 34,013 ops/sec ±0.38% (100 runs sampled)
Fastest is encode (proposal)

benchmark.js

const Benchmark = require('benchmark');
const latest = require('./latest');
const proposal = require('./');
const map = JSON.parse(require("fs").readFileSync(`dist/sourcemap-codec.umd.js.map`));
const mappings = map.mappings;
const decoded = latest.decode(mappings);

console.log(process.version);

const decode = new Benchmark.Suite;
decode
	.add('decode (latest)', () => {
		latest.decode(mappings);
	})
	.add('decode (proposal)', () => {
		proposal.decode(mappings);
	})
	// add listeners
	.on('cycle', (event) => {
		console.log(String(event.target));
	})
	.on('complete', function () {
		console.log('Fastest is ' + this.filter('fastest').map('name'));
	})
	.run({});

console.log('');

const encode = new Benchmark.Suite;
encode
	.add('encode (latest)', () => {
		latest.encode(decoded);
	})
	.add('encode (proposal)', () => {
		proposal.encode(decoded);
	})
	// add listeners
	.on('cycle', (event) => {
		console.log(String(event.target));
	})
	.on('complete', function () {
		console.log('Fastest is ' + this.filter('fastest').map('name'));
	})
	.run({});

[jridgewell]

Friendly ping @Rich-Harris, is there anything you'd like me to do with this PR?