Change Request: `Program` range cover whole program

[overlookmotel]

Which packages would you like to change?

• espree
• eslint-scope
• eslint-visitor-keys

What problem do you want to solve?

The source range for Program is surprising when source code starts or ends with comments.

e.g.:

// foo

x

// bar

Espree (like Esprima) produces an AST where range of Program node excludes whitespace and comments from start and end of file. In example above, the range of Program node is [ 8, 9 ] i.e. the range of the sole statement.

AST explorer - Espree

This differs from Acorn, which produces a range of [0, 17] i.e. the whole of the source code.

AST explorer - Acorn

What is particularly surprising is that with option comment: true you get comments which are outside the range of the program that contains them:

{
  type: 'Program',
  start: 8,
  end: 9,
  body: [
    {
      type: 'ExpressionStatement',
      start: 8,
      end: 9,
      expression: { type: 'Identifier', start: 8, end: 9, name: 'x' }
    }
  ],
  sourceType: 'script',
  comments: [
    { type: 'Line', value: ' foo', start: 0, end: 6, range: [ 0, 6 ] },
    { type: 'Line', value: ' bar', start: 11, end: 17, range: [ 11, 17 ] }
  ]
}

As a further surprise, if the file contains only comments, then the behavior changes:

// foo

This produces a range of [ 0, 8 ]. This does not align with Esprima, which produces range of [ 8, 8 ].

What do you think is the correct solution?

Espree's docs state that:

The primary goal is to produce the exact same AST structure and tokens as Esprima, and that takes precedence over anything else.

Given that Esprima has not received updates for 7 years, and is essentially defunct, I'm not sure whether the maintainers of Espree consider alignment with Esprima still to be the primary goal. I'm also not sure whether that aim, if it still holds, extends to source ranges, or just the AST shape.

What's the correct behavior?

This is arguable. But, in my opinion, it's unintuitive that comments which are part of a program exist in ranges outside of the Program.

The example where the program consists of only comments I think is illustrative. What's the logical range for Program to have in this case, if comments are excluded? [0, 0] or [8, 8]?

The fact that there isn't a clear answer I feel reveals there's a problem with the mental model.

[0, 8] (as Espree currently does) makes most sense to me. But this is inconsistent with Esprima, and also inconsistent with Espree's behavior when there are statements in the program.

Others may feel differently.

Why does this matter?

In addition to the difference from Acorn, @typescript-eslint/parser is also different again - for the first example, it gives a range of [8, 18]! Though the maintainers seem to think Acorn's behavior would be preferable.

AST explorer - TS-ESLint

Personally, I think it's unfortunate that the 3 most depended-on ESTree-compatible parsers Espree, Acorn, and TS-ESLint differ in this respect.

What to do?

One way or another, in my opinion Espree's current behavior is incorrect.

• In the case which program is only comments, output differs from Esprima, violating Espree's stated goal.
• Or, if alignment with Esprima is not an aim as far as ranges are concerned, it would make more sense to have a less surprising behavior.

Changing behavior to match Acorn (which personally I'd advocate for) would be a breaking change. Though, given that the 3 most popular ESTree-compatible parsers differ from each in this regard and I'm not aware of anyone complaining about it before, probably not a big one!

Participation

• I am willing to submit a pull request for this change.

Additional comments

No response

[fisker]

Duplicate of #488

[nzakas]

Thanks for the feedback. As I stated in #488 (comment), Espree is the canonical ESLint parser. Any parser wanting to work with ESLint should mimic what Espree does. As such, we need to be careful when making changes.

The decision of where to start the Program node has long since passed. It will remain as it was with Esprima.

I think producing comments that fall outside of the range of Program is probably a bug, as ESTree is meant to have each node wholly contain its children. @eslint/eslint-tsc what do you think?

For reference, here's the code that adjusts the Program location:

js/packages/espree/lib/espree.js

Lines 184 to 210 in 2800ec8

	/*
	* Adjust opening and closing position of program to match Esprima.
	* Acorn always starts programs at range 0 whereas Esprima starts at the
	* first AST node's start (the only real difference is when there's leading
	* whitespace or leading comments). Acorn also counts trailing whitespace
	* as part of the program whereas Esprima only counts up to the last token.
	*/
	if (program.body.length) {
	const [firstNode] = program.body;
	if (program.range) {
	program.range[0] = firstNode.range[0];
	}
	if (program.loc) {
	program.loc.start = firstNode.loc.start;
	}
	program.start = firstNode.start;
	}
	if (extra.lastToken) {
	if (program.range) {
	program.range[1] = extra.lastToken.range[1];
	}
	if (program.loc) {
	program.loc.end = extra.lastToken.loc.end;
	}
	program.end = extra.lastToken.end;
	}

I think the issue is that comments are not represented as tokens, so the end location currently only considers non-comment syntax.

[overlookmotel]

Thanks very much for coming back. Sorry that I didn't see this had been raised before in #488.

Thanks for confirming that the primary goal for Espree remains aligning with Esprima. In that context, of course the answer is clear. The current behavior cannot change.

Only 2 questions remain:

Incompatibility with Esprima

If the program contains only comments e.g.:

// foo

Espree produces a range of [ 0, 8 ]. This does not align with Esprima, which produces range of [ 8, 8 ].

I assume that divergence from Esprima is unintended? If so, should that be considered a bug (fixing it not a breaking change)? Or something to tackle in next major release?

Comments outside of Program

ESTree is meant to have each node wholly contain its children

Personally I don't think that needs to be considered to apply to comments, as comments are not "nodes" in the sense of the ESTree spec, which doesn't say anything about comments (same as tokens are not "nodes").

Often legal comments or "directive" / "pragma" comments appear at top of file, and it'd likely be problematic for users if those comments were omitted.

Espree's NPM docs specifically state that hashbangs are supported. Hashbangs are treated as comments, and by definition must be at top of file. So excluding all comments that lie outside the range of Program would break that support.

[nzakas]

I assume that divergence from Esprima is unintended? If so, should that be considered a bug (fixing it not a breaking change)? Or something to tackle in next major release?

I think this was unintended as I can't imagine we tested this case when we switched to Acorn. I think this is an indicator that we just need to come to a decision on one question: are comments considered part of Program or not? If yes, then the current behavior for a file with just a comment is correct and we need to adjust what happens when there are comments at the end of a program to make it align. If no, then we should fix the behavior of the one-comment case and leave the other case alone.

[overlookmotel]

are comments considered part of Program or not?

In my view they are. But you've clearly stated that the goal is to align with Esprima, so I thought that answered the question - Esprima does not consider comments to be part of the Program.

[nzakas]

@overlookmotel As I noted, there is an inconsistency in the implementation. Trailing comments are not treated as part of the program but leading comments are. I'm seeking to resolve that inconsistency.

[mdjermanovic]

@nzakas I don't see an inconsistency between how leading comments are treated vs how trailing comments are treated.

I believe Program's range currently works as follows:

1. If there are no tokens, i.e., the source code is only whitespace and/or comments, then the Program's range is the whole source code.
2. If there's at least one token, then the Program's range starts from the first token's start and ends at the last token's end. Whitespace and comments before the first token and after the last token are not included in the Program's range.

Is that correct? If not, can you provide an example where there's an inconsistency between how leading and trailing comments are treated?

[nzakas]

Sorry, I don't think I explained the inconsistency I'm seeing very well. Let me try again.

For instance, here's an example with just a comment:
https://astexplorer.net/#/gist/7896dbd46cd027a8414f45d6c5c5e763/055c0d167a560b52c8cc48a614aacfb50ef07906

In this case, the Program range is 0 to 181.

If I add some code underneath it, the range changes to 178 to 182:
https://astexplorer.net/#/gist/4acf3a930a735d191dd9891280a0a7b9/d3d4bc7af2a20c476676b9f4fc502714eebf6e63

That's the inconsistency I'm concerned with. It seems like the start of the program should be the start of the program. (Similarly, the end of the program should be the end of the program.) But we are counting leading and trailing comments differently in these situations.

[overlookmotel]

Thanks for coming back. I believe the only change required to resolve this inconsistency is as follows:

• If the file contains no tokens (no statements), set Program range start and end both to the length of the source text.

i.e. in your first example, Program's range would be [181, 181].

This would match Esprima's behavior.

There are other alternative solutions which could also resolve the inconsistency, but they would not align with Esprima, which I've understood from your previous comments is of paramount importance.

I hope I've made sense this time!

[mdjermanovic]

For instance, here's an example with just a comment: https://astexplorer.net/#/gist/7896dbd46cd027a8414f45d6c5c5e763/055c0d167a560b52c8cc48a614aacfb50ef07906

In this case, the Program range is 0 to 181.

If I add some code underneath it, the range changes to 178 to 182: https://astexplorer.net/#/gist/4acf3a930a735d191dd9891280a0a7b9/d3d4bc7af2a20c476676b9f4fc502714eebf6e63

That's the inconsistency I'm concerned with. It seems like the start of the program should be the start of the program. (Similarly, the end of the program should be the end of the program.) But we are counting leading and trailing comments differently in these situations.

I still don't see an inconsistency between leading and trailing comments. If there are no tokens, the range spans from the start to the end of the source code, whether the source code content is whitespace-only or contains one or multiple comments. If there is at least one token, the range spans from the start of the first token to the end of the last token, and neither leading nor trailing comments are included.

For example, in the following case the Program range is [0, 56]. The whole source code has 56 code units, and all of the leading comment, trailing comment and whitespace are included in the range:

/**
 * Leading comment
 */

/**
 * Trailing comment
 */

https://astexplorer.net/#/gist/675eeef0df636f1a0297bd48acb29f74/f8e6977b4274852d115199907142b967b1a6e081

While in the following case the Program range is [28, 31]. It's the range of foo, and neither leading nor trailing comments are included:

/**
 * Leading comment
 */

foo

/**
 * Trailing comment
 */

https://astexplorer.net/#/gist/a0cc72f7d660b457ef68bbea44e6eae4/4f83a3bba3a633e3d03a56b5c2b8c4583e9bcf6e