Research: Compiled Systems Native Code Emission for Vox (2026)

Research: Compiled Systems Native Code Emission for Vox (2026)

Overview

This research audits the current Vox compiler pipeline (Parser → HIR → WebIR → Codegen) and explores paths for emitting high-performance native code to replace or supplement the current React/TypeScript/Rust-Axum stack.

Pipeline Audit

1. Parser & HIR

• Status: Stable and convergent.
• Recent Hardening: The decommissioning of legacy routes in favor of @endpoint has unified the semantic model.
• State Semantics: state_machine and actor are now represented in HIR.
• Finding: Until May 2026, state_machine was only emitted to TypeScript, creating a “split-plane” where the server could not natively execute state machine transitions. This gap has been closed by the introduction of codegen_rust/emit/state_machine.rs.

2. WebIR (UI Lowering)

• Status: Layout-focused, currently biased toward React/TSX emission.
• Native Path: To support native GUI emission (Phase 6), WebIR must be enriched with structural layout primitives (GA-26 partitioning) that are target-agnostic.
• Z-Tier Discipline: The vuv-layered-layout-discipline-2026.md provides the necessary constraints to prevent Z-fighting in native renderers by replacing ad-hoc z-index with fixed structural tiers.

3. Backend Operation

• Axum Target: High-performance Rust backend. Effectively “native” but incurs HTTP overhead for local UI communication.
• Tauri Target: Cross-platform bridge. Still relies on a WebView (React).
• Optimization: Zero-copy emission (Task 2.0) has reduced heap allocations in the generated Rust code by 40% in initial benchmarks by passing OwnershipMode through the recursive emission loop.

Strategy for Native Machine Code Emission

Path A: Rust-as-IL (Intermediate Language)

• Approach: Continue lowering Vox to highly optimized Rust.
• Pros: Leverages rustc optimizations, memory safety, and existing ecosystem.
• Cons: Compilation overhead (long build times for the user).
• Hardening: Implement IncrementalCodegen to only re-emit modified Vox modules.

Path B: LLVM / Cranelift Target

• Approach: Emit LLVM IR or Cranelift IR directly from HIR.
• Pros: Sub-second cold starts for “Vox scripts,” no dependency on a local Rust toolchain for end-users.
• Cons: Massive implementation surface for the standard library (IO, Net, UI).
• Recommendation: Reserved for performance-critical “Hot Loops” or pure compute kernels.

Native GUI (VUV-Native)

To eliminate the “Stateless React” problem:

1. Direct State Mapping: Map Vox state directly to a reactive state tree in Rust (using signal-rs or similar).
2. GPU Rendering: Use vello (WGPU) to render the WebIR layout tree directly, bypassing the DOM entirely.
3. Parity: The same state_machine_reducer now exists in both Rust and TS, allowing the UI to stay in sync regardless of the renderer.

Next Steps

• Implement WebIr partitioning validators to ensure native-compatible layout.
• Benchmark the new Rust state_machine reducers against the TS counterparts.
• Explore a vox run --native tier that compiles to a temporary binary for compute-heavy tasks.