Skip to content

0010-100-year-language.md

Latest commit

 

History

History
224 lines (173 loc) · 11.1 KB

0010-100-year-language.md

File metadata and controls

224 lines (173 loc) · 11.1 KB
Raw

RFC 0010: 100-Year Language Architecture

  • Status: Draft
  • Tier: N/A (design guidance, not surface change)
  • Author: David Porkka (AI-assisted)
  • Created: 2026-07-25
  • Resolved: TBD
  • Language-version at effect: N/A
  • Supersedes: none
  • Superseded by: none

Summary

Documents architectural decisions and design principles that position Nulang for relevance across a 100+ year horizon — spanning Von Neumann, non-Von Neumann, quantum, and event-driven architectures, with small binaries, powerful libraries, and no AI terminology lock-in. This RFC does not propose surface changes; it records design rationale and recommends migration paths for existing surfaces that couple the language to current paradigms.

Motivation

Programming languages that survive decades do so by separating invariant semantics from transient implementation details. C's longevity comes from its thin abstraction over the machine; Erlang's from the actor model; Lisp's from its minimal core. Nulang already has three strengths that map to longevity: the actor model (concurrency primitive that predates multi-core and will outlive it), algebraic effects (a generalization of exceptions, async, and generators that composes cleanly), and a frozen Core subset (RFC 0002). This RFC captures the remaining architectural decisions needed.

Design

C.1 Frozen Core as the Longevity Contract

The Core subset defined in RFC 0002 is the invariant kernel that every conforming Nulang implementation must support. Its longevity properties:

  • Self-hosting path: The bootstrap compiler (bootstrap/compiler_core.nula) is written in Core and targets the .nbc format (RFC 0001). Stage 1 compiles Core programs; Stage 2 compiles itself. This means Nulang can be ported to new hardware without the Rust toolchain — only a Core interpreter or compiler is needed, and the bootstrap compiler provides it.

  • .nbc format stability (RFC 0001): Frozen bytecode artifacts carry a version number, BLAKE3 source hash, and language version. The format migration registry (src/format/migrate.rs) is the sole legal home for format upgrades. Old artifacts remain runnable as long as the migration path exists.

  • Value layout versioning: The VALUE_LAYOUT_VERSION constant gates the i64-tagged representation. A future version can support 128-bit or variable-width tags without breaking the frozen Core semantics — only the runtime encoding changes.

  • Wire protocol versioning: The NUL0 wire protocol (magic NUL0, 8-byte node-id handshake, versioned Packet enum) carries a protocol version that gates compatibility. All distribution primitives are behind the frozen Actor.* and Otp.* built-in effect interfaces — the transport can evolve independently.

Recommendation: These mechanisms already exist. No change required; this section documents them as the longevity contract.

C.2 Decouple AI Terminology from Language Surface

Current AI surface couples the language to the LLM paradigm:

Current name Problem Recommendation
LLM effect (LLM.ask) "Large Language Model" is a specific architecture Rename to Inference effect (Inference.ask) — provider-agnostic, covers any ML inference
@tool annotation Good as generic annotation; semantics are extensible Keep as-is; document extensibility
Pipeline module Hardcoded orchestration pattern Move to standard library; language should not hardcode orchestration
Supervisor module Same Move to standard library
Debate module Same Move to standard library
agent keyword Already unified with actor in v2.0 design (SPEC2.md Forward) Complete the migration: agent → actor alias, deprecate agent

Migration path:

  1. LLM → Inference: Add Inference as an alias; deprecate LLM over one major version; remove in the following version.
  2. Pipeline, Supervisor, Debate: Extract from built-in modules to a nulang-ai standard library package. The language surface should expose only actor, perform Inference.ask(...), and @tool.
  3. agent → actor: Already aligned in SPEC2.md v2.0. Complete the parser desugaring so agent becomes a deprecated alias for actor.

C.3 Architecture-Independent Value Representation

Current i64-tagged value layout (48-bit payload, 16-bit tag) assumes 64-bit architectures. For multi-architecture relevance:

  • Path to 128-bit: The value layout version number allows a future layout with 112-bit payload and 16-bit tag, or a variable-width encoding. The Core semantics (integers are arbitrary-precision in the type system, bounded by i64 at runtime) already distinguish logical from physical representation.

  • Endianness: The .nbc format and NUL0 wire protocol use big-endian encoding. Future non-little-endian hardware (some quantum control processors, neuromorphic chips) would need explicit endianness negotiation. Add an endianness flag to the NUL0 handshake (NUL0 magic + version + endianness byte) in a future protocol version.

  • Quantum computing path: Values are classical; quantum operations would be a new effect (Quantum) with separate state. The effect system already isolates effects — a Quantum effect would carry its own qubit register state, invisible to classical code. No change to Core needed.

  • Tagged vs. uniform: The Core type system (HM + capabilities + effect rows) is representation-agnostic. A future implementation could use uniform boxed values or a different tagging scheme without changing any Core program.

Recommendation: No surface change. Add endianness negotiation to the NUL0 protocol in a future protocol version. Document the value layout version as the migration mechanism.

C.4 Non-Von Neumann and Event-Driven Architecture Path

The actor model + algebraic effects maps naturally to diverse execution models:

  • Event-driven architectures: Actors are event handlers. Messages are events. The mailbox is an event queue. The receive expression with selective matching (ReceiveMatch opcode) is event pattern-matching. No change needed — the actor model was designed for this.

  • Dataflow / non-Von Neumann: Effect handlers can be compiled to dataflow graphs. A perform is a node activation; the continuation is a dataflow edge; resume is backpressure resolution. The MIR→bytecode pipeline already has a lowering phase (src/mir_lower.rs). A MIR→dataflow backend would be a new codegen target (src/mir_dataflow.rs), not a language change.

  • Spatial / reconfigurable (FPGA, CGRAs): The MIR is a control-flow graph with basic blocks. A MIR→spatial backend would map basic blocks to reconfigurable logic regions, with perform/handle as region transitions. Again, a codegen target, not a language change.

  • Neuromorphic: The actor model's message-passing maps to spiking neural networks: actors are neurons, messages are spikes, mailboxes are synaptic delays. The Inference effect (renamed from LLM, see C.2) would be the natural interface to neuromorphic accelerators.

Recommendation: Document these mappings. No surface change needed — the existing architecture already decomposes cleanly across execution models. The MIR layer is the right abstraction boundary for new backends.

C.5 Small Binaries + Powerful Libraries

Current: single binary includes VM, JIT, WASM, AI, Python, LSP. For small binaries and library distribution:

  • Feature flags already exist: --no-default-features strips Python, SQLite, LSP, and AI runtime. Document minimal build sizes per feature set. A bytecode-only build (no JIT, no WASM, no native, no Python, no LSP, no AI) should fit in ~1 MB.

  • Pre-compiled library artifacts: The nula package manager should support .nbc artifacts as library dependencies. A library author publishes type-checked, compiled .nbc files; consumers link them without source distribution. The .nbc format already carries type metadata (via the constant pool and function signatures) — extend it with an export table for library symbols.

  • Bootstrap compiler artifacts: The bootstrap compiler produces .nbc artifacts that can be as small as hundreds of bytes for simple programs. This is the path to tiny standalone binaries: compile to .nbc, embed a minimal Core VM (~50 KB), ship as a single executable.

Recommendation: Document minimal build configurations. Extend the nula package manager with .nbc library dependency support as a follow-up RFC.

C.6 Syntax Stability

The keyword set in SPEC2.md §2.3 lists 57 keywords. Every reserved word is a permanent tax on the namespace.

  • Freeze the current set as the "Frozen Syntax" tier. Any new keyword must go through the RFC process and be gated by a language version.

  • Unwired reserved words: The following keywords are reserved but not wired into the implementation: priv, loop, node, monitor, link, exit, await, subworkflow. These should either be wired (with RFCs) or removed from the lexer before 2.0. Reserved-but-unused keywords create confusion and block user identifiers for no benefit.

  • Deprecation path: A keyword removed from Frozen tier must:

    1. Be marked deprecated in the lexer (warning on use as keyword, allowed as identifier with a migration warning)
    2. Remain reserved for one major version
    3. Be freed in the following major version
    4. Migration tooling (nulang migrate) rewrites affected source files

Recommendation: Before 2.0, audit the 57 keywords and either wire or remove the 8 unwired ones. Document the keyword lifecycle in GOVERNANCE.md.

Tier Classification

This RFC does not change any stability tier. The recommendations affect:

  • Frozen: Value layout versioning, .nbc format, NUL0 protocol (already frozen). The endianness flag would be a backward-compatible extension.
  • Stable: The LLM → Inference rename would be a Stable-tier change (deprecation cycle required).
  • Experimental: Pipeline, Supervisor, Debate extraction to standard library is Experimental-tier (can change without deprecation).

Backwards Compatibility

All recommendations are backward-compatible in their initial phase:

  • LLM → Inference: Add alias, deprecate old name, remove later.
  • agent → actor: Parser desugaring, no existing code breaks.
  • Pipeline/Supervisor/Debate extraction: Keep existing imports working while adding standard library equivalents.
  • Keyword removal: Deprecation cycle with migration tooling.

References

  • RFC 0001: Format Stability (.nbc format, migration registry)
  • RFC 0002: Frozen Core (Core subset, bootstrap compiler)
  • SPEC2.md §1.1a: Nulang Core
  • SPEC2.md §2.3: Keywords
  • src/value_layout.rs: Value layout constants and versioning
  • src/format/: Format stability layer (.nbc artifacts, migration registry, NUL0 wire protocol)
  • src/runtime/network.rs: NUL0 wire protocol (handshake, Packet enum)
  • src/stdlib.rs: Built-in effect inventory (IO, LLM, Timer, Signal, Actor, Otp)
  • AGENTS.md: Architecture overview, pipeline, value layout, distribution