lyng/notes/object_indexer_extension_optimization.md

Object indexer extension follow-up

Context

• Obj.getAt / Obj.putAt now support scoped extension indexers for named singleton object declarations, so patterns like Storage["name"] can route through override fun Storage.getAt(...) and override fun Storage.putAt(...).
• Dispatch order is intentionally:
  1. concrete non-root class members
  2. scoped extensions
  3. built-in string-field fallback

Why this order

• Normal class-defined indexers are the common fast path and should not pay extension lookup cost first.
• Checking scoped extensions earlier caused avoidable overhead for regular overridden indexers.
• Routing extension indexers through the generic root Obj.getAt member path caused recursion; direct extension dispatch avoids that.

Possible future optimization

• If extension-based indexers show up in hot profiles, add a tiny receiver-shape/member cache at the extension-dispatch points in Obj.getAt / Obj.putAt.
• The cache should stay behind the concrete non-root member fast path.
• A good shape key is likely (receiver class id/layout version, member name) with the cached extension record/wrapper.