Generator performance

[JLHwung]

Q	A
Fixed Issues?
Patch: Bug Fix?
Major: Breaking Change?
Minor: New Feature?
Tests Added + Pass?	Yes
Documentation PR Link
Any Dependency Changes?
License	MIT

This PR improves the @babel/generator performance.

• The Buffer storage backend is switched to string, it was a micro-optimization in 6 yrs ago. But now the string approach (more intuitive) is faster according to the benchmark
• The popular path Buffer#_endsWith are rewritten to avoid scanning the _queue
• String comparison like string[i] === "another one byte string" is replaced by string.charCodeAt(i) === charCodes
• Access the "isQueries" in @babel/types is extremely slow. In this PR we hoist them to top level. Unfortunately we can not use the import namespace specifiers such as import { isIdentifier } from "@babel/types" because Babel replaces them to types.isIdentifier.

See commit messages for commit-by-commit performance reports.

In a real world case jquery-3.6.txt, we see 50% performance improvements compared to @babel/generator 7.14.5

baseline 1 jquery 3.6: 11.74 ops/sec ±12.2% (85ms)
baseline 2 jquery 3.6: 6.89 ops/sec ±5.4% (145ms)
current 1 jquery 3.6: 19.97 ops/sec ±9.56% (50ms)
current 2 jquery 3.6: 10.27 ops/sec ±4.38% (97ms)

Many empty statements

baseline 256 empty statements: 3_455 ops/sec ±18.71% (0.289ms)
baseline 512 empty statements: 2_029 ops/sec ±1.29% (0.493ms)
baseline 1024 empty statements: 1_006 ops/sec ±1.37% (0.994ms)
baseline 2048 empty statements: 482 ops/sec ±2.01% (2.074ms)
current 256 empty statements: 5_334 ops/sec ±22.11% (0.187ms)
current 512 empty statements: 3_085 ops/sec ±1.73% (0.324ms)
current 1024 empty statements: 1_541 ops/sec ±1.73% (0.649ms)
current 2048 empty statements: 751 ops/sec ±1.95% (1.332ms)

Many 1-length identifiers

baseline 256 1-length identifiers: 1_686 ops/sec ±30.51% (0.593ms)
baseline 512 1-length identifiers: 1_079 ops/sec ±1.38% (0.927ms)
baseline 1024 1-length identifiers: 533 ops/sec ±1.64% (1.876ms)
baseline 2048 1-length identifiers: 267 ops/sec ±1.67% (3.749ms)
current 256 1-length identifiers: 2_788 ops/sec ±25.92% (0.359ms)
current 512 1-length identifiers: 1_587 ops/sec ±1.34% (0.63ms)
current 1024 1-length identifiers: 792 ops/sec ±1.58% (1.262ms)
current 2048 1-length identifiers: 365 ops/sec ±3.4% (2.742ms)

Many 25-length identifiers

baseline 256 25-length identifiers: 1_792 ops/sec ±25.73% (0.558ms)
baseline 512 25-length identifiers: 984 ops/sec ±2.46% (1.016ms)
baseline 1024 25-length identifiers: 520 ops/sec ±2.24% (1.924ms)
baseline 2048 25-length identifiers: 254 ops/sec ±1.91% (3.932ms)
current 256 25-length identifiers: 2_435 ops/sec ±49.59% (0.411ms)
current 512 25-length identifiers: 1_555 ops/sec ±1.78% (0.643ms)
current 1024 25-length identifiers: 751 ops/sec ±1.94% (1.331ms)
current 2048 25-length identifiers: 363 ops/sec ±1.65% (2.753ms)

[codesandbox-ci]

This pull request is automatically built and testable in CodeSandbox.

To see build info of the built libraries, click here or the icon next to each commit SHA.

Latest deployment of this branch, based on commit 9321162:

Sandbox	Source
babel-repl-custom-plugin	Configuration
babel-plugin-multi-config	Configuration

[babel-bot]

Build successful! You can test your changes in the REPL here: https://babeljs.io/repl/build/47494/

[nicolo-ribaudo]

Access the "isQueries" in @babel/types is extremely slow. In this PR we hoist them to top level. Unfortunately we can not use the import namespace specifiers such as import { isIdentifier } from "@babel/types" because Babel replaces them to types.isIdentifier.

We might use a custom plugin to replace import { isIdentifier } from "@babel/types" to import * as _t from "@babel/types"; const { isIdentifier } = _t in every package.

[jridgewell]

Nit: It would probably make sense to run this as a babel transformer automatically? That would let us take care of charCodes.foo, too.

[nicolo-ribaudo]

The transform-charcodes plugin replaces things like charCodes.lineFeed with 10 so it's even better!