Future Features

Planned features and future work for the Ashes language and ecosystem. Shipped features are documented in the normative docs under docs/md/ — syntax/semantics in Language Spec, library APIs in Standard Library, runtime/backend behavior in Architecture, and the history of the compiler's optimization/codegen work in the Compiler Changelog — not here.

Feature	Status	Description
Self-Hosting	Exploratory	Rewrite the compiler in Ashes
Self-Hosting Log	Record	The completed work of that rewrite
WebAssembly Target	Exploratory	A wasm32 backend for browsers and sandboxed plugin hosts
Package Registry Website: Follow-up Work	Deferred product work	Follow-up work for the shipped public package-discovery website
Compiler Parity Fuzzing	Proposed	Differentially fuzz the stage-0 and self-hosted compilers against each other
Inferred Borrow Lifetimes for Ordinary Values	Research proposal	Investigate richer borrow-lifetime inference to remove provably unnecessary RC operations
Optimizer Pass Observability	Proposed	Optional per-pass instrumentation to see which optimizer pass changed what

---

Ground Rules

1. Spec first. Update the Language Reference before implementing any new syntax or semantic rule.
2. Layer discipline. Respect the project dependency graph (Frontend → Semantics → Backend). Runtime behaviour never goes in Frontend.
3. Test every invariant. Each feature must ship with tests that prove the new guarantees.
4. No user-visible Drop. Drop is a compiler concept. Users see automatic cleanup.
5. Purity preserved. All values are immutable. There is no mutation. All APIs — standard library and user-defined — are pure: they return new values and never modify their arguments. There are no in-place updates visible to user code.
6. No tracing GC. Ordinary lifetime operations are compiler-inserted; resource cleanup remains statically verified.

---

Remaining Optimizer Opportunities

The 2026-08 optimizer audit shipped its backlog; what it built and measured is recorded in the Compiler Changelog, the current pipeline in Architecture, and the self-hosting deltas in Self-Hosting. These are the pieces it deliberately left open, with the finding that should gate any future attempt so it is not re-derived:

• Match compilation beyond one grouping level. Arms are grouped by outer constructor tag with one shared tag test per group; a multi-case group still re-tests the outer tag per case, and there is no column reordering or guard interaction within a group. Any dead-arm elimination must use the sound recursive coverage engine (TryGetMissingPatternCore, the one behind the "missing case" diagnostic), never top-level tag coverage — tag coverage was tried twice and is unsound for nested sub-patterns (Ok(true) | Ok(false) | Error(_)) — and must unify each guard-free pattern's type with the scrutinee first, because a parameter's type is routinely still unresolved when the match is lowered.
• Tail contification of local helpers. Deferred. TcoContext is the backbone of the enclosing function's per-parameter RC/arena representation decisions, so a second join point per contified helper means a second instance of that machinery inside one frame. The motivating case also disappeared: once a real read-builtin RC leak found during baselining was fixed, a local helper and its hand-inlined form measured identically at -O0 and -O2. Find a program where they differ before implementing.
• Multi-anchor Perceus drop placement. Implemented, then reverted on evidence. Placing one drop for a value with several lexical anchors needs a union of reachable regions that is unsound across a TCO back-edge (a HasBackEdge guard fixes the segfault), and after the fix every real and synthetic program produced byte-identical IR: lowering already places each branch's drop at that branch's true last use. Revisit only with a program where a lexical anchor provably sits later than the branch's real last use.
• Closure devirtualization to a small label set. A call whose closure can be one of 2-4 known labels could dispatch directly per arm (lambda-set specialization); today only a single agreeing label devirtualizes.
• Closures with three or more scalar captures. One capture rides the env word and a second the otherwise-unused ownership-flag word of the shared three-word call signature; a third needs a direct-call-only worker with a per-function parameter list. A worker for a unary function alone is what LLVM's dead-argument elimination already does at -O2, so build the convention only together with a measured consumer (N>=3 scalarization, uncurrying, or contification of curried helpers).
• Open-world reuse past a statically-resolved callee. In-place reuse borrowing now survives a hand-off to a callee proven inspect-only by a whole-program fixpoint; a higher-order or unresolved callee still forces the defensive copy.
• Mutual-recursion merging for slot types without a default. Groups whose differing parameter types are user-declared, tuples, functions, or unresolved keep the closure path; merging them needs the loop's per-parameter active-flag machinery to tolerate an uninitialized inactive slot rather than a default literal.
• Heap-closure local helpers. A let-bound helper that also escapes keeps its runtime-managed closure and RcDrop; only stack closures lose their environment through scalarization.
• Interprocedural summary unification. Only the fixpoint skeleton (WholeProgramFixpoint) is shared; FunctionOwnershipSummary (AST-phase, FuncKey-keyed) and the IR-phase label-keyed analyses stay separate by design, since forcing one node type across that phase boundary buys nothing their consumers need.