Appearance

0010. Independent codegen backends — no shared crate ​

Date: 2026-05-29

Status ​

Accepted

Context ​

neoc-codegen-lua and neoc-codegen-js carry structurally-identical pieces: the EnumIndex / TupleStructIndex use-site resolution tables, and the pattern-lowering Compiled shape in each crate's pat.rs. The bodies differ only in target syntax (and vs &&, local vs let, [1] vs ._0).

The DRY pressure is to extract the shared logic into a common crate both backends depend on. Two forces push back:

  • Crate isolation. The workspace enforces per-crate boundaries in Cargo.toml, not convention. A neoc-codegen-common crate adds a new node to the graph that both backends must track; hoisting the pieces into neoc-ast widens the frozen AST contract's responsibility beyond pure data.
  • Backends diverge by design. Lua and JS are different targets with different runtime models (per-value metatables vs a prototype registry). The shared-looking pieces are shared today; coupling them through a dependency bets that they stay identical, and pays an indirection cost at every site if they don't.

The duplication is small (two index structs and one pattern module) and the backends are each well under a maintainable size.

Decision ​

The two codegen backends are kept independent. The duplicated index and pattern-lowering logic is accepted rather than extracted into a shared crate, and is not hoisted into neoc-ast. Each backend owns its full lowering and may diverge as its target requires. The intentional parallel is documented on both EnumIndex structs, and the two are kept in sync by hand.

Consequences ​

  • Easier: each backend evolves freely per target with no shared dependency to coordinate; crate isolation is preserved; no premature abstraction over a similarity that may not hold.
  • Harder: the index/pattern logic is duplicated — a bug fixed in one backend must be mirrored in the other, and the hand-sync expectation relies on discipline (and the documenting comments), not the compiler.
  • Revisit when: the backends prove the shared pieces are stable across a few language-feature additions. If the duplication stops drifting, the cost/benefit of a shared crate flips and this can be superseded.