What

check_function_redeclaration (the Annex B.3.3 legacy duplicate-function rule) distinguished a plain FunctionDeclaration from an async/generator one by reading the previous declaration's AST node through its stored NodeId:

let prev_function = ctx.nodes.kind(prev.declaration).as_function();
if prev_function.is_some_and(|f| !(f.r#async || f.generator)) { return; }

That node is a popped sibling (not an ancestor of the current node), so it's the last build-time read that would prevent making AST node storage optional.

How

The async/generator shape is now recorded at bind time, where the node is current and free to read off the Function (self.r#async || self.generator):

• Function::bind passes it down via a new declare_function_symbol.
• Each Redeclaration record carries is_async_or_generator. When the first declaration's record is reconstructed on the first redeclaration, its value comes from build-only SemanticBuilder scratch (an IndexVec<SymbolId, bool> populated at symbol creation and discarded after the build) — not from the node.
• check_function_redeclaration reads prev.flags.contains(Function) && !prev.is_async_or_generator — no node access.

Notes

• The scratch is bind-time-only: it never reaches the transformer and isn't persisted on Scoping, so there's nothing to keep in sync — no SymbolFlags bit and no transform_checker mask are needed (redeclarations are compared by span only).
• Behaviour is unchanged — conformance is zero snapshot diff.
• check_redeclaration's named-function-expression read is intentionally left on the node: that node is the enclosing function (an ancestor), so it stays reachable in parent-pointer mode.

Prerequisite for making AST node storage optional (#22859).

🤖 Generated with Claude Code