Track AST node between passes

[Boshen]

In Rolldown, ast node information needs to be tracked across scans and transforms.

It's currently using spans to track the nodes, but once transformation is in, newly inserted statements does not have a SPAN so information is lost, and Rolldown will not function anymore :-(

---

From hyf: how to identity the same ast node in different visit passes

There's workaround for rolldown to use faked span that out of orignal text's length.
But this require oxc could handle span that is out of original source's length

[overlookmotel]

As discussed in meet just now, I can see solutions to this which doesn't require AST node IDs.

1. Record "path" to node

e.g. body[3] or body[5].expression.left.

For import / export statements, recording the "path" is simple as they're always top-level statements. So all you need is the statement index. For import.meta or dynamic import(), it's less good as these expressions can be anywhere in the AST so the path can be deep. But still probably faster than searching the entire AST for a Span.

NB: "path" would not be stored as a string, but as a Vec<PathSegment> where PathSegment is an enum. Something like:

struct Path(Vec<PathSegment>);

enum PathSegment {
  ProgramBody(u32),
  ProgramDirective(u32),
  ProgramHashtag,
  ExpressionStatementExpression,
  BinaryExpressionLeft,
  BinaryExpressionRight,
  /* ... lots more ... */
}

fn statement_from_path_mut<'a, 'p>(program: &'p mut Program<'a>, path: &Path) -> &'p mut Statement<'a> {
  todo!();
}

fn expression_from_path_mut<'a, 'p>(program: &'p mut Program<'a>, path: &Path) -> &'p mut Expression<'a> {
  todo!();
}

2. Record *mut pointer to node

This would be faster as you can jump direct to the node without any searching/traversing the AST. But would need quite a bit of thought to make a safe API.

[Boshen]

Our semantic pass already generates an ast node id in AstNode 🤔

[overlookmotel]

Yes, but that part of semantic is not usable when the AST is mutable, due to aliasing restrictions. This is why Traverse runs semantic, but then throws away the part of Semantic which contains AstNodes - borrow checker won't allow keeping it.

[Boshen]

Tracking in oxc-project/oxc#4188 for research on AST Node ID.