Unambiguous JavaScript Grammar

Status	DRAFT
Authors	@jdalton @bmeck
Date	June 14, 2016

Proposal Accepted!

On July 6, 2016 this proposal was accepted as section 5.1 of the Node ES6 Module Interoperability draft proposal ‼️

TL;DR

• CJS and ES modules just work without new extensions, extra ceremony, or excessive scaffolding
• Performance is generally on par with existing CJS module loading
• Performance is significantly improved for ES modules over transpilation workflows
• Change JS grammars for Script and Module to be unambiguous
• Determine grammar of .js files by parsing as one grammar and falling back to others

Problem

The Script and Module goal of ECMA262 have a grammatical ambiguity where some code can run in both goals, having the exact same source text, but produce different results. Unlike "use strict", the signal to have a specific behavior is not in the code, thus the code has a multitude of possible effects which are not controlled by the programmer.

Note: The following example highlights a few of the effects of ambiguous grammar and is by no means exhaustive.

Example

function foo(value) {
  value = value || '';
  var args = [].slice.call(arguments);
  args.unshift(this);
  return args;
}
foo(null);

Differences:

	script	module
variable scope of foo	global	local
arguments object	modified	unmodified
this binding of foo	global	undefined

Since there is no way in source text to enforce the goal with the current grammar; this leads to the behavior of certain constructs being undefined by the programmer, and defined by the host environment. In turn, existing code could be run in the wrong goal and partially function, or function without errors but produce incorrect results.

ECMA262 Solution

Require that Module source text has at least one import or export declaration. A module with only an import declaration and no export declaration is valid. Modules, that do not export anything, should specify an export {} to make intentions clear and avoid accidental parse errors while removing import declarations. The export {} is not new syntax and does not export an empty object. It is simply the standard way to specify exporting nothing.

Note: While the ES2015 specification does not forbid this extension, Node wants to avoid acting as a rogue agent. Node has a TC39 representative, @bmeck, to champion this proposal. A specification change or at least an official endorsement of this Node proposal would be welcomed. If a resolution is not possible, this proposal will fallback to the previous .mjs file extension proposal.

Script Example

function foo(value) {
  value = value || '';
  var args = [].slice.call(arguments);
  args.unshift(this);
  return args;
}
foo(null);

	script	module (cannot parse)
variable scope of foo	global	n/a
arguments object	modified	n/a
this binding of foo	global	n/a

Module Example

function foo(value) {
  value = value || '';
  var args = [].slice.call(arguments);
  args.unshift(this);
  return args;
}
foo(null);
export {};

	script (cannot parse)	module
variable scope of foo	n/a	local
arguments object	n/a	unmodified
this binding of foo	n/a	undefined

Problem

Node currently requires a means for programmers to signal what goal their code is written to run in.

Leading solutions have either hefty ecosystem tolls, ceremony, or scaffolding. They lack a way to define the intent of the source text from the ECMA262 standard.

Solution

A package opts-in to the Module goal by specifying "module" as the parse goal field (name not final) in its package.json. Package dependencies are not affected by the opt-in and may be a mix of CJS and ES module packages. If a parse goal is not specified, then attempt to parse source text as the preferred goal (Script for now since most modules are CJS). If there is a parse error that may allow another goal to parse, then parse as the other goal, and so on. After this, the goal is known unambiguously and the environment can safely perform initialization without the possibility of the source text being run in the wrong goal.

Algorithm

Parse (source, goal, throws)

The abstract operation to parse source text as a given goal.

1. Bootstrap source for goal.
2. Parse source as goal.
3. If success, return true.
4. If throws, throw exception.
5. Return false.

Operation

1. If a package parse goal is specified, then

2. Let goal be the resolved parse goal.

3. Call Parse(source, goal, true) and return.

4. Else fallback to multiple parse.

5. If Parse(Source, Script, false) is true, then 1. Return.

6. Else 1. Call Parse(Source, Module, true).

Note: A host can choose either goal to parse first and may change their order over time or as new parse goals are introduced. Feel free to swap the order of Script and Module.

Implementation

To improve performance, host environments may want to specify a goal to parse first. This can be done in several ways:
cache on disk, a command line flag, a manifest file, HTTP header, file extension, etc.

Tooling Concerns

Some tools, outside of Node, may not have access to a JS parser (Bash programs, some asset pipelines, etc.). These tools generally operate on files as opaque blobs / plain text files and can use the techniques, listed under Implementation, to get parse goal information.

External Examples and Impact

• Esprima relies on a --module flag to signal the Module goal. However, this has proven to be unintuitive for many users. Unambiguous Script and Module goals would enable things to “just work” without flags.

• Facebook Flow performs a series of inferences to detect CJS and ES modules. Unambiguous Script and Module goals would improve its ability to determine module types.

• JSCS can accept input through stdin, so identification of parse goals by source text is ideal.

• Linters, like xo, could use unambiguous Script and Module goals to enable module specific linting rules without extra configuration.

• Microsoft packaged web applications can benefit from unambiguous Script and Module goals. The bytecode cache for a packaged web application is generated upon installation. When the application is running, files are loaded by script tags so their intended parse goals are understood. However, bytecode cache generation is done without running the application so the intended parse goals are unknown. Because of this, the bytecode cache is generated for the Script goal and ignored for ES modules.

• TypeScript has taken the stance from early on that a script becomes a module when it has at least one import or export declaration. Over the years they have experienced very few user issues with this approach.

Special Thanks

This proposal would not have been possible without the tireless effort, conviction, and collaboration of
@bmeck, @dherman, and @wycats.

While iterating on this proposal we have reached out to several people from areas affected by it.
Although opinions on this proposal are varied, I am grateful for feedback from:

• @AtOMiCNebula (Microsoft Edge)
• @brendaneich (TC39)
• @bterlson (Microsoft Chakra / TC39)
• @chrisdickinson (Node)
• @hzoo (Babel)
• @indutny (Node)
• @leobalter (jQuery Foundation / TC39)
• @ljharb (TC39)
• @loganfsmyth (Babel)
• @mathiasbynens (Lodash)
• @rvagg (Node)
• @sheerun (Bower)
• @sindresorhus (AVA / Chalk / Yeoman)
• @travisleithead (Microsoft Edge / W3C)
• @trevnorris (Node)

Thank you! ❤️
— @jdalton