Type table emits unreachable types and members (O(N^2)) and carries vestigial member-dedup infrastructure

[zharinov]

Problem

The emitted type table carries dead weight. Neither issue affects correctness. They are pure binary bloat plus leftover code.

(a) Unreachable type defs/members are serialized into the binary.

TypeTableBuilder::build does two passes:

1. Collect reachable types from iter_def_types().
2. Then unconditionally sweep every interned type via iter_types().

The second sweep drags in every intermediate "progressive-merge" struct that inference created. There is no type DCE. The number of dead members grows O(N^2) with alternation width and nesting depth.

# An 8-branch unlabeled alternation: 14 dead type_defs + 35 dead members; >80% of the type table is dead.
cargo run -p plotnik -- dump -l javascript -q 'Q = [(identifier) @a (number) @b (string) @c (array) @d (object) @e (true) @f (false) @g (null) @h]'

The live members sit at the lowest indices (the named/result type is collected first), so the materializer never reads a wrong index. This is bloat, not miscompilation.

(b) Vestigial member-dedup infrastructure.

• member_cache (emit/type_table.rs:34) is built and read but never written, so it is always empty.
• MemberRef::Deferred is never constructed by production code. Every production site uses deferred_by_index.
• crates/plotnik-compiler/src/bytecode/ir.rs:141 has .expect("deferred member reference must resolve"). This is a latent panic. It fires the day someone wires up the deferred dedup that the compile/scope.rs:45 comment promises but does not populate the cache.
• Two comments are misleading:
  • scope.rs:45 says "will be added later" for dedup that was never added.
  • emit/emitter.rs:202 says "struct fields are deduplicated by field identity". That is false; members are emitted per-struct with duplication (e.g. a: T16 repeats across many structs in the dump above).

Cause

• (a) emit/type_table.rs:64-72 (the unconditional second sweep), analyze/type_check/context.rs:217-222 (iter_types returns all interned types), and no type DCE in compile/dce.rs (it prunes instructions only).
• (b) member_cache at emit/type_table.rs:34/:45/:350; MemberRef::Deferred resolve at crates/plotnik-compiler/src/bytecode/ir.rs:140-141; misleading comments at scope.rs:45 and emitter.rs:202.

Fix

• (a) After build, mark-and-sweep type_defs reachable from entrypoint result types plus TypeName entries, remap member_start, drop the rest. Or drop the iter_types() sweep if the enum-inside-named-node edge it guards is reachable through iter_def_types.
• (b) Pick one:
  • Remove member_cache, TypeTableBuilder::lookup_member, the emitter.rs:205 closure, and MemberRef::Deferred/deferred().
  • Or implement the deferred dedup and actually populate the cache.
  • Either way, fix or delete the two misleading comments.

Acceptance

• dump of an N-branch unlabeled alternation shows no unreferenced type_defs/members.
• No dead member_cache/Deferred surface remains, or the latent .expect is provably unreachable by construction.

Related

The latent .expect panic in (b) blocks the deferred "call-site member dedup" feature that the scope.rs:45 comment defers. That feature cannot be built via MemberRef::deferred() until member_cache is populated, or compilation aborts. The O(N^2) dead members in (a) compound with the merge-alternation null-fill growth.