Adds Babel plugin babel-plugin-optimize-react

[trueadm]

This PR adds a Babel 7 plugin that aims to optimize certain React patterns that aren't as optimized as they might be. For example, with this plugin the following output is optimized as shown:

// Originalvar_useState=Object(react__WEBPACK_IMPORTED_MODULE_1_["useState"])(Math.random()),_State2=Object(_Users_gaearon_p_create_rreact_app_node_modules_babel_runtime_helpers_esm_sliceToArray_WEBPACK_IMPORTED_MODULE_0__["default"])(_useState,1),value=_useState2[0];// With this pluginvaruseState=react__WEBPACK_IMPORTED_MODULE_1_.useState;var__ref__0=useState(Math.random());varvalue=__ref__0[0];

Named imports for React get transformed

// OriginalimportReact,{useState}from'react';// With this pluginimportReactfrom'react';const{useState}=React;

Array destructuring transform for React's built-in hooks

// Originalconst[counter,setCounter]=useState(0);// With this pluginconst__ref__0=useState(0);constcounter=__ref__0[0];constsetCounter=__ref__0[1];

React.createElement becomes a hoisted constant

// OriginalimportReactfrom'react';functionMyComponent(){returnReact.createElement('div',null,'Hello world');}// With this pluginimportReactfrom'react';const__reactCreateElement__=React.createElement;functionMyComponent(){return__reactCreateElement__('div',null,'Hello world');}

[apostolos]

It doesn't pick up the correct reference to React when default import is not used (e.g. non-JSX modules).

Let's say you have the following module:

import{useRef,useEffect}from'react';exportfunctionusePrevious<T>(value: T){constref: React.MutableRefObject<T|null>=useRef<T>(null);useEffect(()=>{ref.current=value;});returnref.current;}

Current output is:

// EXTERNAL MODULE: ./node_modules/react/index.jsvarreact=__webpack_require__(602);varreact_default=/*#__PURE__*/__webpack_require__.n(react);// ...var_React=React,useRef=_React.useRef,useEffect=_React.useEffect;functionusePrevious(value){varref=useRef(null);useEffect(function(){ref.current=value;});returnref.current;}

Expected output:

// EXTERNAL MODULE: ./node_modules/react/index.jsvarreact=__webpack_require__(602);varreact_default=/*#__PURE__*/__webpack_require__.n(react);// ...varuseRef=react_default.useRef,useEffect=react_default.useEffect;functionusePrevious(value){varref=useRef(null);useEffect(function(){ref.current=value;});returnref.current;}

[trueadm]

@apostolos I've published a new version to NPM – let me know if that fixes your issue.

[apostolos]

@trueadm 0.0.2 fixed the issue with the default import, thanks!

I've found one more issue, although minor. The following is currently broken:

importReactfrom'react';import{memo}from'react';

I get the following error:

ModuleParseError: Module parse failed: Identifier 'React' has already been declared (3:15)
You may need an appropriate loader to handle this file type.
| import React from 'react';| var __reactCreateElement__ = React.createElement;> import { memo, React } from 'react';

The question is why bother since you can import both in one statement. It will probably cause issue with TypeScript users that don't use allowSyntheticDefaultImports. The following style is not uncommon:

import*asReactfrom'react';import{useRef,useEffect,memo}from'react';

[gaearon]

We definitely want to support all kinds of imports, thanks for reporting.

[gaearon]

(We should also check that const React = require('react') and similar are at least not broken)

[trueadm]

@apostolos Thanks for reporting, I'll fix that.
@gaearon There are tests for CJS requires. :)

Update: fix published, you can find it on NPM now.

[Andarist]

this is gonna load babelrc / babel.config.js, not sure if you want that here

also - would be good to have tests both with & without @babel/plugin-transform-destructuring (and possibly some other overlapping combinations)

[Andarist]

I see you have no babelrc files in this package, but probably better disable config loading with

babel.transform(code,{babelrc: false,configFile: false,plugins: [plugin],}).code

just to be more future proof

[eps1lon]

Good call on testing with other plugins. Destructuring optimization seems to be ignored when used alongside preset-env: optimize-react + preset-env.

I'll work on a PR to this one. At least with a test, hopefully with a fix.

[Andarist]

this optimization is OK, but it works on per module (file) basis - it doesnt take into account that production bundles use scope hoisting, maybe we could somehow end up with single constant reference per chunk instead of many?

[trueadm]

That would be good, but I was unsure how to do that?

[Jessidhia]

Don't hardcode the identifier name, use path.scope.generateUidIdentifier in Program and save its value on state (the second argument to a visitor, which also is === this)

[Jessidhia]

I guess having a hardcoded name actually does let what Andarist suggested work, but not in a way that works with a minifier. You'd have to do var ref = ref || React.createElement over and over and you can't prove to a minifier the result of that expression is guaranteed truthy. Scope hoisting would rename the local vars anyway.

The only surefire way to get this working with scope hoisting is to have an ESM export of React. It is possible to generate one with a webpack plugin, and then use a Babel plugin to rewrite imports of React to import the generated ESM wrapper. (e.g. rewrite "from 'react'" to be "from 'react-esm-loader!react'")

This way, you could ensure that only other ESM would be importing the shared React ESM, and scope hoisting would do its job.

That'd stop working as well if there are dynamic imports, though, as webpack would be forced to put the modules shared with more than one chunk in a separate, non-scope-hoisted module. Perhaps another webpack plugin pass could detect this case and generate one wrapper module per chunk.

[gaearon]

We don't enable scope hoisting in CRA. Tbh I think it's reasonable compromise for now.

[Andarist]

We don't enable scope hoisting in CRA. Tbh I think it's reasonable compromise for now.

Are you sure? I haven't checked it, but it's enabled by default in webpack@4 in production mode (which is used by CRA). https://webpack.js.org/plugins/module-concatenation-plugin/

That would be good, but I was unsure how to do that?

Me neither 😅 Would have to move this rewrite to other phase than transpilation.

[apostolos]

I think I've found an edge case: This module defines an FC that passes its children prop to createPortal, it does not render JSX directly.

Source: https://github.com/LWJGL/lwjgl3-www/blob/e0ee6d0d47d92c453c1803c6a8b4b0c2ed63a156/client/components/Portal.tsx

The import looked like this:

import{memo,useRef,useEffect}from'react';

But it breaks like in the first case (_React instead of react_default).

I found the following 3 workarounds:

// 1importReact,{useRef,useEffect}from'react';const{ memo }=React;exportconstPortal=memo(/*...*/);// 2importReact,{useRef,useEffect}from'react';exportconstPortal=React.memo(/*...*/);// 3importReact,{memo,useRef,useEffect}from'react';exportconstPortal=memo(/*...render at least some JSX here...*/);

Important detail: If memo is not used at all, it works fine!

[vincentriemer]

Gave it a try and it appears not to be working with namespace imports.

I cloned your branch and added this test:

it('should transform React.createElement calls #4',()=>{consttest=` import * as React from "react"; const node = React.createElement("div", null, React.createElement("span", null, "Hello world!")); export function MyComponent() { return node; } `;constoutput=transform(test);expect(output).toMatchSnapshot();});

Which results in the following snapshot:

exports[`React createElement transforms should transform React.createElement calls #4 1`]=`"import * as React, React from \\"react\\";const node = React.createElement(\\"div\\", null, React.createElement(\\"span\\", null, \\"Hello world!\\"));export function MyComponent() { return node;}"`;

Which explains the syntax errors I was getting.

[trueadm]

@apostolos@vincentriemer Thanks for the bug reports. I'll fix them tomorrow. If you want to get involved though – feel free to make a PR against my forked React repro. Any help would be grateful :)

[artembatura]

@trueadm Hey, nice work! What about these Babel plugins? It's relevant now?

[trueadm]

@artemirq Many of those plugins are still relevant but not in the scope right now for this plugin. We may expand this scope in the future.

[trueadm]

@vincentriemer@apostolos I've released a new version of the plugin to NPM. Please let me know if you find anymore issues. Thanks for the great help!

[apostolos]

@trueadm Fixes all remaining issues for me (also tested import * as React from 'react'). Using it in production here: https://www.lwjgl.org/

[trueadm]

@apostolos Awesome stuff. Did you notice and differences compared to before (bundle size, performance)?

[apostolos]

Did a quick comparison on lwjgl.org (latest React alpha, uses only function components with hooks, no classes). I've included react-local which does similar optimizations (although, afaik, without the hook transformation):

All routes with content

Build	Minified	Minified/GZIP
baseline	1004.45 kB	275.91 kB
react-local	1004.13 kB	276.26 kB
optimize-react	1004.13 kB	276.25 kB

/customize route (more app-like)

Build	Minified	Minified/GZIP
baseline	66.66 kB	21.39 kB
react-local	66.53 kB	21.55 kB
optimize-react	66.53 kB	21.55 kB

---

• Both minify better but compress worse than baseline.
• I expected more dramatic differences when compiled with spec: true, loose: false, but didn't see any. It might be because Babel supposedly now assumes array when compiling hooks destructuring... I don't know.
• Didn't see noticeable performance difference between them and profiling didn't help. I suppose it depends on the app and I'm sure we'll see something on an artificial benchmark.

EDIT: Source code available here: https://github.com/LWJGL/lwjgl3-www/tree/master/client

[trueadm]

@apostolos Thanks for checking. :) The differences are all very negligible indeed!

[gaearon]

To be fair that example uses TS so I'm not sure it's directly comparable to our current setup.

[chrisvasz]

This is great, but doesn't appear to play nicely with @babel/preset-env.

babel.config.js:

module.exports={presets: ['@babel/react',['@babel/preset-env',{modules: false,targets: 'ie>=11'}],],plugins: ['optimize-react'],};

App.js:

importReact,{useState}from'react';functionApp(){let[count,setCount]=useState(0);return<divonClick={()=>setCount(count+1)}>{count}</div>;}

output from npx babel App.js (whether or not targets: 'ie>=11' is included):

function_slicedToArray(arr,i){return_arrayWithHoles(arr)||_iterableToArrayLimit(arr,i)||_nonIterableRest();}function_nonIterableRest(){thrownewTypeError("Invalid attempt to destructure non-iterable instance");}function_iterableToArrayLimit(arr,i){var_arr=[];var_n=true;var_d=false;var_e=undefined;try{for(var_i=arr[Symbol.iterator](),_s;!(_n=(_s=_i.next()).done);_n=true){_arr.push(_s.value);if(i&&_arr.length===i)break;}}catch(err){_d=true;_e=err;}finally{try{if(!_n&&_i["return"]!=null)_i["return"]();}finally{if(_d)throw_e;}}return_arr;}function_arrayWithHoles(arr){if(Array.isArray(arr))returnarr;}importReactfrom'react';var__reactCreateElement__=React.createElement;varuseState=React.useState;functionApp(){var_useState=useState(0),_useState2=_slicedToArray(_useState,2),count=_useState2[0],setCount=_useState2[1];return__reactCreateElement__("div",{onClick: functiononClick(){returnsetCount(count+1);}},count);}

In this output, it doesn't look like the array destructuring transform happened. If I comment out the line in babel.config.js that includes @babel/preset-env, the output looks much better:

importReactfrom'react';const__reactCreateElement__=React.createElement;const{
useState
}=React;functionApp(){let_ref_0=useState(0);letsetCount=_ref_0[1];letcount=_ref_0[0];return__reactCreateElement__("div",{onClick: ()=>setCount(count+1)},count);}

[avocadowastaken]

Mostly functional components, tons of hooks. Parsed size become lower, gzipped is bigger:

Without `babel-plugin-optimize-react`

Parsed

Show chunks:
All (1.45 MB)
static/js/vendors~main.3aebfb9f.chunk.js (591.12 KB)
static/js/main.f1f1f4a0.chunk.js (416.35 KB)
static/js/vendors~admin.8473e02e.chunk.js (340.38 KB)
static/js/admin.0ebde3f8.chunk.js (135.28 KB)
static/js/bundle.2f4bc937.js (2.24 KB)

Gzipped

All (382.77 KB)
static/js/vendors~main.3aebfb9f.chunk.js (175.99 KB)
static/js/main.f1f1f4a0.chunk.js (106.06 KB)
static/js/vendors~admin.8473e02e.chunk.js (79.7 KB)
static/js/admin.0ebde3f8.chunk.js (19.87 KB)
static/js/bundle.2f4bc937.js (1.15 KB)

With `babel-plugin-optimize-react`

Parsed

Show chunks:
All (1.42 MB)
static/js/vendors~main.3aebfb9f.chunk.js (591.12 KB)
static/js/main.65336000.chunk.js (397.01 KB)
static/js/vendors~admin.8473e02e.chunk.js (340.38 KB)
static/js/admin.3364ead4.chunk.js (126.3 KB)
static/js/bundle.94a853c1.js (2.24 KB)

Gzipped

All (384.5 KB)
static/js/vendors~main.3aebfb9f.chunk.js (175.99 KB)
static/js/main.65336000.chunk.js (107.03 KB)
static/js/vendors~admin.8473e02e.chunk.js (79.7 KB)
static/js/admin.3364ead4.chunk.js (20.63 KB)
static/js/bundle.94a853c1.js (1.15 KB)

[trueadm]

@umidbekkarimov Overall that looks to be a general positive win. Gzip size was marginally down but parsed time is where it really went up. Also Brotli compression should further improve over gzip in the cases where the plugin was enabled.

[eps1lon]

This should be configurable or ideally communicate with targets/browserslists/@babel/preset-env. Array spread syntax is supported by all evergreen browsers and that for quite some time now. The optimization is only interesting if you need to support IE11. See example preset-env without IE11

[trueadm]

We opted to do this because the runtime performance of spread syntax was considerably slower in all browsers when I tests this compared to the transformed version. This might have changed since, as this plugin was created 7 months ago.

[eps1lon]

Makes sense. Not sure what the current status of destructuring optimizations in v8 but they had plans to improve it a few months ago. Do you have some numbers about the performance gain?

[eps1lon]

@material-ui/core: parsed: -0.53%, gzip: +0.77%

-- mui/material-ui#16072 (comment)

With brotli -q 11 it's 72_080 (w/o) vs. 72_683.

[Friss]

@trueadm Wanted to check in on the status of this getting into CRA. Its certainly an interesting optimization.

Also wanted to make a note where we ran into an issue with the usage of a lowercase react import.

import react from 'react';
import propTypes from 'prop-types';
import createReactClass from 'create-react-class';

transforms to

import React, react from 'react';
const __reactCreateElement__ = React.createElement;
import propTypes from 'prop-types';
import createReactClass from 'create-react-class';

[trueadm]

@Friss I don't believe this will be going in. The gains weren't really there and there wasn't much appetite from folks either.

[gaearon]

Maybe we can just get in the getter fix. https://twitter.com/sebmarkbage/status/1250284377138802689?s=21