Skip to content

Latest commit



467 lines (330 loc) · 37.6 KB

File metadata and controls

467 lines (330 loc) · 37.6 KB

core-js@3, babel and a look into the future

After more than 1.5 years of development, dozens of pre-releases, many sleepless nights, core-js@3 is finally released. It's the largest set of changes in core-js and polyfilling-related babel features of all time.

What is core-js?

  • It is a polyfill of the JavaScript standard library, which supports:
    • The latest ECMAScript standard.
    • ECMAScript standard library proposals.
    • Some WHATWG / W3C standards (cross-platform or closely related ECMAScript).
  • It is maximally modular: you can easily choose to load only the features you will be using.
  • It can be used without polluting the global namespace.
  • It is tightly integrated with babel: this allows many optimizations of core-js import.

It's the most universal and the most popular way to polyfill JavaScript standard library, but a big part of developers just don't know that they use core-js indirectly 🙂


core-js is my own hobby project and it does not bring me any profit. It takes too much time and it is really costly: to finish work on core-js@3, I had left my job some months ago. This project facilitates the life of many people and companies. For these reasons, it makes sense to start raising funds to support the maintenance of core-js.

If you are interested in the core-js project or use it in your day-to-day work, you can become a sponsor on Open Collective or Patreon.

You can propose me a good job where I will be able to work on something related.

Or you can contribute in another way: you can help improve code, tests or documentation (currently, core-js documentation is terrible!).

What changed in core-js@3?

Changes in JavaScript standard library content

Because of the following two reasons, this release is rich with new JavaScript polyfills:

  • core-js only has breaking changes in major releases, even if it is needed to reflect a change in a proposal.
  • core-js@2 entered feature freeze 1.5 years ago; all new features were added only to the core-js@3 branch.

Stable ECMAScript features

Stable ECMAScript features had already been almost completely supported by core-js for a long time, however, core-js@3 introduced some new features:

Some features that have already been available for a long time as proposals have been accepted in ES2016-ES2019 and are now marked as stable:

Added many fixes for browsers bugs/issues. For example, Safari 12.0 Array.prototype.reverse bug has been fixed.

ECMAScript proposals

In addition to supported before, core-js@3 now supports the following ECMAScript proposals:

  • globalThis stage 3 proposal - before, we had global and
  • Promise.allSettled stage 2 proposal
  • New Set methods stage 2 proposal:
    • Set.prototype.difference
    • Set.prototype.intersection
    • Set.prototype.isDisjointFrom
    • Set.prototype.isSubsetOf
    • Set.prototype.isSupersetOf
    • Set.prototype.symmetricDifference
    • Set.prototype.union
  • New collections methods stage 1 proposal, which includes many new useful methods:
    • Map.groupBy
    • Map.keyBy
    • Map.prototype.deleteAll
    • Map.prototype.every
    • Map.prototype.filter
    • Map.prototype.find
    • Map.prototype.findKey
    • Map.prototype.includes
    • Map.prototype.keyOf
    • Map.prototype.mapKeys
    • Map.prototype.mapValues
    • Map.prototype.merge
    • Map.prototype.reduce
    • Map.prototype.some
    • Map.prototype.update
    • Set.prototype.addAll
    • Set.prototype.deleteAll
    • Set.prototype.every
    • Set.prototype.filter
    • Set.prototype.find
    • Set.prototype.join
    • Set.prototype.reduce
    • Set.prototype.some
    • WeakMap.prototype.deleteAll
    • WeakSet.prototype.addAll
    • WeakSet.prototype.deleteAll
  • String.prototype.replaceAll stage 1 proposal
  • String.prototype.codePoints stage 1 proposal
  • Array.prototype.last(Item|Index) stage 1 proposal
  • compositeKey and compositeSymbol methods stage 1 proposal
  • Number.fromString stage 1 proposal
  • Math.seededPRNG stage 1 proposal
  • Promise.any (with AggregateError) stage 0 proposal

Some proposals have been largely changed, and core-js was updated accordingly:

Web standards

Many useful features have been added to this category.

The most important one is support for URL and URLSearchParams. It was one of the most popular feature requests. Adding URL and URLSearchParams, making maximally spec-compliant, supporting any environment keeping their source code small compact was one of the hardest tasks in the core-js@3 development.

core-js@3 includes a standard method to create microtasks in JavaScript: queueMicrotask. core-js@2 provided the asap function which did the same thing and was an old ECMAScript proposal. queueMicrotask is defined in the HTML standard and it is already available in modern engines like Chromium or NodeJS.

Another popular feature request was support for the .forEach method on DOM collection. Since core-js already polyfilled iterators of DOM collections, why not add also .forEach to NodeList and DOMTokenList?

Removed obsolete features:

  • Reflect.enumerate because it's removed from the spec
  • and global since now they are replaced by globalThis
  • Array.prototype.flatten since it's replaced by Array.prototype.flat
  • asap since it's replaced by queueMicrotask
  • Error.isError has withdrawn a long time ago
  • RegExp.escape rejected a long time ago
  • Map.prototype.toJSON and Set.prototype.toJSON also rejected a long time ago
  • Unnecessary CSSRuleList, MediaList, StyleSheetList iteration methods which were added mistakenly

No more non-standard non-proposed features

Many years ago, I started writing a library which I needed as the core of my JavaScript applications: this library contained polyfills and some utilities for common needs. After some time, it was published as core-js. I think that at this moment most core-js users do not use non-standard core-js features. Almost all of them were removed in previous releases, and it's time to remove all the remaining ones from core-js. Starting from this release, core-js can be finally called a polyfill.

Packages, entry points and modules names

A popular issue was a big size (~2MB) of the core-js package and duplication of many of its files. For this reason, core-js was split into three packages:

  • core-js, which defines global polyfills. (~500KB, 40KB minified and gzipped)
  • core-js-pure, which provides polyfills without pollution the global environment. It's the equivalent of core-js/library from core-js@2. (~440KB)
  • core-js-bundle: a bundled version of core-js which defines global polyfills.

In previous versions of core-js, modules with polyfills for stable ECMAScript features and ECMAScript proposals were prefixed with es6. and es7. respectively. It was a decision taken in 2014 when all the features which could be after ES6 were considered as ES7. In core-js@3 all stable ECMAScript features are prefixed with es., while ECMAScript proposals with esnext..

Almost all CommonJS entry points were changed. In core-js@3 there are many more entry points than there were in core-js@2: they bring maximum flexibility, making it possible to include only the polyfills needed by your application.

Here are some examples of how the new entry points can be used:

// polyfill all `core-js` features:
import "core-js";
// polyfill only stable `core-js` features - ES and web standards:
import "core-js/stable";
// polyfill only stable ES features:
import "core-js/es";

// if you want to polyfill `Set`:
// all `Set`-related features, with ES proposals:
import "core-js/features/set";
// stable required for `Set` ES features and features from web standards
// (DOM collections iterator in this case):
import "core-js/stable/set";
// only stable ES features required for `Set`:
import "core-js/es/set";
// the same without global namespace pollution:
import Set from "core-js-pure/features/set";
import Set from "core-js-pure/stable/set";
import Set from "core-js-pure/es/set";

// if you want to polyfill just required methods:
import "core-js/features/set/intersection";
import "core-js/stable/queue-microtask";
import "core-js/es/array/from";

// polyfill reflect metadata proposal:
import "core-js/proposals/reflect-metadata";
// polyfill all stage 2+ proposals:
import "core-js/stage/2";

Some other important changes

It's now possible to configure the aggressiveness of core-js polyfills. If you think that core-js feature detection is too aggressive in some cases and that the native implementation is correct enough for your use-case, or if an incorrect implementation isn't detected by core-js as such, you can change the core-js default behavior.

If a feature can't be implemented following the specification in every detail, core-js adds a .sham property to the polyfill. For example, in IE11 Symbol.sham is true.

No more LiveScript! When I started the core-js project, I mainly used LiveScript; after some time, I rewrote all the polyfills in JavaScript. Tests and helper tools in core-js@2 still used LiveScript: it is a very interesting CoffeeScript-like language with powerful syntax sugar which allows writing very compact code, but now it's almost dead. Other than that, it was an additional barrier for contributing to core-js because most core-js users do not know this language. core-js@3 tests and tools use modern ES syntax: it could be a good moment to start contributing to core-js 🙂

For almost all users, for optimization of core-js import, I recommend using babel. However, core-js-builder is still useful in some cases. Now it supports the targets argument which takes a browserslist query with target engines - you can create a bundle which contains only required for target engines polyfills. For cases like this, I made the core-js-compat package, more info about it you could find in @babel/preset-env part of this article.

This is just the tip of the iceberg, much more changes were done internally. You can find more info about core-js changes in the changelog.


As mentioned above, babel and core-js are tightly integrated: babel gives the possibility of optimizing the core-js import as much as possible. A serious part of work on core-js@3 was improving core-js-related babel features (see this PR). Those changes are published in Babel 7.4.0.


@babel/polyfill is a wrapper package which only includes imports of stable core-js features (in Babel 6 it also included proposals) and regenerator-runtime/runtime, needed by transpiled generators and async functions. This package doesn't make it possible to provide a smooth migration path from core-js@2 to core-js@3: for this reason, it was decided to deprecate @babel/polyfill in favor of separate inclusion of required parts of core-js and regenerator-runtime.

Instead of

import "@babel/polyfill";

you should use those 2 lines:

import "core-js/stable";
import "regenerator-runtime/runtime";

Don't forget install those dependencies directly!

npm i --save core-js regenerator-runtime


@babel/preset-env has 2 different modes, which can be enabled with the useBuiltIns option: entry and usage, which optimize imports of core-js in different ways.

Babel 7.4.0 introduces both changes commons to the two modes and specific to each mode.

Since @babel/preset-env now supports core-js@2 and core-js@3, useBuiltIns requires setting a new option, corejs, which specifies the used version (corejs: 2 or corejs: 3). If it isn't directly set, corejs: 2 will be used by default and it will show a warning.

To make it possible for Babel to support new core-js features introduced in future minor versions, you also can specify the minor core-js version used in your project. For example, if you want to use core-js@3.1 and take advantage of new features added in that version, you can set the corejs option to 3.1: corejs: '3.1' or corejs: { version: '3.1' }.

One of the most important parts of @babel/preset-env was the source providing data about the features supported by different target engines, to understand whether something needs to be polyfilled by core-js or not. caniuse, mdn and compat-table are good educational resources but aren't really meant to be used as data sources for developer tools: only the compat-table contains a good set of ES-related data and it is used by @babel/preset-env, but it has some limitations:

  • it contains data only about ECMAScript features and proposals, but not about web platform features like setImmediate or DOM collections iterators. So, up to now, @babel/preset-env added all web platform features from core-js even for targets where they are supported.
  • it does not contain any information about (even serious) bugs in engines: for example, already mentioned Array#reverse broken in Safari 12 but it isn't marked as unsupported by compat-table. On the other hand, core-js correctly fixes broken implementations, but with compat-table this capability wasn't taken advantage of.
  • it contains only some basic and naive tests, which do not check that features work as they should in real-world cases. For example, old Safari has broken iterators without .next method, but compat-table shows them as supported because it just checks that typeof of methods which should return iterators is "function". Some features like typed arrays are almost completely not covered.
  • compat-table is not designed for providing data for tools. I'm one of the compat-table maintainers, but some of the other maintainers are against maintaining this functionality.

For this reason, I created the core-js-compat package: it provides data about the necessity of core-js modules for different target engines. When using core-js@3, @babel/preset-env will use that new package instead of compat-table. Please help us with testing and providing data and mappings for missing engines! 😊

Until Babel 7.3, @babel/preset-env had some problems related to the order polyfills were injected. Starting from version 7.4.0, @babel/preset-env will add the polyfills only when it knows which of them is required and in the recommended order.

useBuiltIns: entry with corejs: 3

When using this option, @babel/preset-env replaces direct imports of core-js with imports of only the specific modules required for a target environment.

Before those changes, @babel/preset-env replaced only import '@babel/polyfill' and import 'core-js', they were synonyms and used for polyfilling all stable JavaScript features.

Since @babel/polyfill is now deprecated, @babel/preset-env doesn't transpile it when corejs is set to 3.

An equivalent replacement for @babel/polyfill with core-js@3 is

import "core-js/stable";
import "regenerator-runtime/runtime";

When targeting chrome 72, it will be transformed by @babel/preset-env to

import "core-js/modules/es.array.unscopables.flat";
import "core-js/modules/es.array.unscopables.flat-map";
import "core-js/modules/es.object.from-entries";
import "core-js/modules/web.immediate";

when targeting chrome 73 (which completely support ES2019 standard library), it will just become a single smaller import:

import "core-js/modules/web.immediate";

Since now @babel/polyfill is deprecated in favor of separate core-js and regenerator-runtime inclusion, we can optimize regenerator-runtime import. For this reason, regenerator-runtime import will be removed from the source code when targeting browsers that support generators natively.

Now, @babel/preset-env in useBuiltIns: entry mode transpile all available core-js entry points and their combinations. This means that you can customize it as much as you want, by using different core-js entry points, and it will be optimized for your target environment.

For example, when targeting chrome 72,

import "core-js/es";
import "core-js/proposals/set-methods";
import "core-js/features/set/map";

will be replaced with

import "core-js/modules/es.array.unscopables.flat";
import "core-js/modules/es.array.unscopables.flat-map";
import "core-js/modules/es.object.from-entries";
import "core-js/modules/esnext.set.difference";
import "core-js/modules/esnext.set.intersection";
import "core-js/modules/";
import "core-js/modules/";
import "core-js/modules/";
import "core-js/modules/";
import "core-js/modules/esnext.set.symmetric-difference";
import "core-js/modules/esnext.set.union";

useBuiltIns: usage with corejs: 3

When using this option, @babel/preset-env adds at the top of each file imports of polyfills only for features used in the current and not supported by target environments.

For example,

const set = new Set([1, 2, 3]);
[1, 2, 3].includes(2);

when targeting an old browser like ie 11, will be transformed to

import "core-js/modules/es.array.includes";
import "core-js/modules/es.array.iterator";
import "core-js/modules/";
import "core-js/modules/es.set";

const set = new Set([1, 2, 3]);
[1, 2, 3].includes(2);

when targeting, for example, chrome 72 no imports will be injected, since those polyfills not required for this target:

const set = new Set([1, 2, 3]);
[1, 2, 3].includes(2);

Until Babel 7.3, useBuiltIns: usage was unstable and not fully reliable: many polyfills were not included, and many others were added without their required dependencies. In Babel 7.4, I tried to make it understand every possible usage pattern.

I improved the techniques used to determine which polyfills should be added on property accesses, object destructuring, in operator, global object property accesses.

@babel/preset-env now injections polyfills required for syntax features: iterators when using for-of, destructuring, spread and yield delegation; promises when using dynamic import, async functions and generators, etc.

Babel 7.4 supports injecting proposals polyfills. By default, @babel/preset-env does not inject them, but you can opt-in using the proposals flag: corejs: { version: 3, proposals: true }.


When used with core-js@3, @babel/transform-runtime now injects polyfills from core-js-pure: a version of core-js that doesn't pollute the global namespace.

core-js@3 and @babel/runtime have been integrated together by adding a corejs: 3 option to @babel/transform-runtime and creating the @babel/runtime-corejs3 package. But what advantages did this bring?

One of the most popular issue with @babel/runtime was that it did not support instance methods. Starting from @babel/runtime-corejs3, this problem has resolved. For example,


will be transpiled to

import _includesInstanceProperty from "@babel/runtime-corejs3/core-js-stable/instance/includes";

_includesInstanceProperty(array).call(array, something);

Another notable change is the support of ECMAScript proposals. By default, @babel/plugin-transform-runtime does not inject polyfills for proposals and use entry points which do not include them but, exactly as you can do in @babel/preset-env, you can set the proposals flag to enable them: corejs: { version: 3, proposals: true }.

Without proposals flag,

new Set([1, 2, 3, 2, 1]);

is transpiled to:

import _Set from "@babel/runtime-corejs3/core-js-stable/set";

new _Set([1, 2, 3, 2, 1]);

when proposals are enabled, it becomes:

import _Set from "@babel/runtime-corejs3/core-js/set";
import _matchAllInstanceProperty from "@babel/runtime-corejs3/core-js/instance/match-all";

new _Set([1, 2, 3, 2, 1]);
_matchAllInstanceProperty(string).call(string, /something/g);

Some other old issues have been fixed. For example, this quite popular pattern didn't work when using @babel/runtime-corejs2 but it is supported with @babel/runtime-corejs3.

myArrayLikeObject[Symbol.iterator] = Array.prototype[Symbol.iterator];

Although previous versions of @babel/runtime did not work with instance methods, iterables (both [Symbol.iterator]() calls and its presence) were supported using some custom helper functions. Extracting the [Symbol.iterator] method was not supported, but now it works.

As a cheap bonus, @babel/runtime now supports IE8-, with some limitations. For example, since IE8- does not support accessors, modules transform should be used in loose mode and regenerator-runtime (which internally uses some ES5+ built-ins) needs to be transpiled by this plugin.

Look into the future

Much work has been done, but core-js is still far from perfect. How can the library and tools be improved in the future and how do language changes can affect it?

Old engines support

At this moment, core-js tries to support all possible engines and platforms where we can test it: it even supports IE8- or, for example, early Firefox versions. While it is useful for some users, only a small part of developers using core-js need it. For many other users, it can cause some problems like bigger bundle size or slower runtime execution.

The main problem comes from supporting ES3 engines (above all, IE8-): most modern ES features are based on ES5 features, which aren't available in those very old browsers.

The biggest missing important feature is property descriptors: when they aren't available, some features can't be polyfilled because they either are accessors (like RegExp.prototype.flags or URL properties setters) or are accessors-based (like typed arrays polyfill). In order to workaround this lack, we need to use different workarounds (for example, to keep Set.prototype.size updated). Maintenance of those workarounds sometimes is too painful, and removing them would highly simplify many polyfills.

However, descriptors are just a part of this problem. The ES5 standard library contains many other features that can be considered as the basis of modern JavaScript: Object.keys, Object.create, Object.getPrototypeOf, Array.prototype.forEach, Function.prototype.bind, etc. Unlike the most modern features, core-js internally relies on them and in order to implement even a simple modern function, core-js needs to load implementations of some of those "building blocks". It is a problem for users who want to create a maximally minimalistic bundle and only import just a few core-js polyfills.

In some countries, IE8 still is quite popular, but browsers should disappear at some point to allow the web to move forward. IE8 was released 19-03-2009; today it is 19-03-2019: it's the 10th birthday of IE8. IE6 is about to turn 18: I stopped testing new core-js versions in IE6 some months ago.

We should drop IE8- and other engines without basic ES5 support in core-js@4.

ECMAScript modules

core-js use CommonJS modules. It has been the most popular JavaScript modules format for a long time, but now ECMAScript provides its own modules format. Many engines already support them; some bundlers (like rollup) are based on them, and some other bundlers provide them as an alternative to CommonJS. It would make sense to provide an alternative version of core-js which uses ECMAScript modules format.

Extended web standards support?

core-js is currently focused on ECMAScript support, but it also supports a few web standards features which are available cross-platform and closely related to ECMAScript. Adding polyfills for web standards like fetch is a very popular feature request.

The main reason why core-js doesn’t include them was that it would have seriously increased bundles size and it would have forced core-js users to load features which might not have been needed. Now core-js is maximally modular, user can include only some chosen features, there are tools like @babel/preset-env and @babel/runtime which helps to get rid of unused or unnecessary polyfills.

Maybe it's time to revisit this old decision?

@babel/runtime for target environment

Currently, we can't set the target environment as @babel/runtime like we can do for @babel/preset-env. That means that @babel/runtime injects all possible polyfills even when targeting modern engines: it unnecessarily increases the size of the final bundle.

Since core-js-compat contains all the necessary data, in the future, it will be possible to add support for compiling for a target environment to @babel/runtime and to add a useBuiltIns: runtime option to @babel/preset-env.

Better optimization of polyfill loading

As explained above, Babel plugins give us different ways of optimizing core-js usage, but they are not perfect: we can improve them.

@babel/preset-env with useBuiltIns: usage now should work much better than before, but it could still fail in some uncommon cases: when the code can't be statically analyzed. For that case, we need to find a way for library developers to specify which polyfills are required by their library instead of directly loading them: some kind of metadata, which will be used to inject polyfills when creating the final bundle.

Another issue of useBuiltIns: usage is the duplication of polyfills import. useBuiltIns: usage can inject dozens of core-js imports in each file. But what if our project has thousands of files or even tenths of thousands? In this case, we will have more lines of code with import "core-js/..." than lines of code in core-js itself: we need a way to collect all imports to one file so that they can be deduplicated.

Almost every @babel/preset-env user which targets old engines like IE11 uses a single bundle for every browser. That means that even modern engines with full ES2019 support will be loading the unnecessary polyfills only required by IE11. Sure, we can create different bundles for different targets and use, for example, the type=module / nomodules attributes: one bundle for modern engines with modules support, another for legacy engines. Unfortunately, it’s not a complete solution to this problem: a service that bundles polyfills for the required target based on the user agent would be really useful. And we already have one - polyfill-service. Although it is an interesting and popular service, polyfills quality leaves much to be desired. It’s not as bad as it was some years ago: the team of this project is actively working to improve it, but I wouldn't recommend using polyfills from this project if you want them to match native implementations. Some years ago was an attempt to use core-js as a polyfills source for this project, but it hadn't been possible because polyfill-service relies on files concatenation instead of modules (like core-js in the first few months after it was published 😊).

A service like this one integrated with a good polyfills source like core-js, which only loads the needed polyfills by statically analyzing the source like Babel's useBuiltIns: usage option does could cause a revolution in the way we think about polyfills.

New features proposals from TC39 and possible problems for core-js

TC39 is working really hard to improve ECMAScript: you can see the progress by looking at all the new proposals implemented in core-js. However, I think that some features of some proposals could cause serious problems for polyfilling / transpiling. There would be enough to say about this topic to write a whole new post, but I'll try to summarize my thoughts here.

Standard library proposal, stage 1

At this moment, TC39 is considering adding to ECMAScript built-in modules: a modular standard library. It would be a great addition to JavaScript, and core-js is the best place where it could be polyfilled. With the techniques used in @babel/preset-env and @babel/runtime, we could theoretically inject polyfills for required built-in modules in a very simple way. However, the current version of this proposal causes some serious problems which don't make it as straightforward.

Polyfilling of built-in modules, as stated by the authors of the proposal, only means falling back to layered APIs or import maps. This means that if a native module will be missing, it will be possible to load a polyfill from a provided URL. That's absolutely not what polyfills need, and it is incompatible with the architecture of core-js and every other popular polyfill project. Import maps shouldn't be the only way to polyfill built-in modules.

We will be able to get a built-in module just by using ES modules syntax with a special prefix. This syntax haven't any equal based on the previous version of the language - transpiled modules will not be able to interact with not transpiled in modern engines - it will cause problems for package distribution.

More other, it will work asynchronously. It's a critical problem for feature detection - scripts will not wait when you'll detect a feature and load a polyfill - feature detection should be done synchronously.

The first implementation of built-in modules without a proper way of transpiling / polyfilling already available. If it will not be revised, built-in modules will not be able to be polyfilled in the current core-js format. The proposed way of polyfilling will seriously complicate the lives of developers.

The issue with the standard library can be solved by adding a new global (maybe it will be the last one?): a registry of built-in modules which will allow getting and setting them synchronously, like

StandardLibraryRegistry.set(moduleName, value);

Asynchronous fallbacks like layered APIs should be used only after this global registry.

As a bonus point, it would simplify transpiling native modules import to old syntax.

Decorators proposal, new iteration, stage 2

In the new iteration of this proposal, it has been seriously reworked. Decorator definitions aren't a syntax sugar anymore and, like with built-in modules, we will not be able to write a decorator in an old version of the language and use it as a native decorator. Other than that, decorators are not just usual identifiers - they live in a parallel lexical scope: this means that transpiled decorators can't interact with native decorators.

The proposal authors recommend distributing packages with untranspiled decorators and leaving to the library consumers the choice to transpile their dependencies. However, it's not possible in different scenarios. This approach could prevent core-js from polyfilling new built-in decorators when they will be added to the JS standard library.

Decorators should be just an alternative way of applying functions on something, they should only be syntax sugar for wrappers. Why complicate things?

If a new language feature does not introduce to the language something fundamentally new, an alternative for what couldn't be implemented in a previous version of the language, we should be able to transpile and/or polyfill it, and transpiled/polyfilled code should be able to interact with the native feature in engines which supports this feature natively.

I hope for the wisdom of the authors of those proposals and of the committee, that these proposals will be adapted so that it will be possible to properly transpile or polyfill them.

If you are interested in the core-js project or use it in your day-to-day work, you can become a sponsor on Open Collective or Patreon. core-js isn't backed by a company: its future depends on you.

Feel free to add comments to this article here.

Denis Pushkarev, 19-03-2019, thanks Nicolò Ribaudo for redaction