Dala Compiler Runtime — Actor-Native Mobile Runtime

A Rust-based actor-native mobile runtime with integrated AI orchestration and typed native execution. Inspired by BEAM semantics but designed from the ground up for mobile constraints: startup time, battery, thermal, memory locality, offline AI, UI synchronization, and native platform integration.

Architecture

Elixir/Erlang
    ↓
BEAM SSA
    ↓
Dala Typed IR  ←─── Stable shapes, message types, tensor types, capabilities
    ↓
Optimization Passes ←─── Pattern matching, mailbox specialization, SIR promotion
    ↓
Backend Lowering
    ├── Interpreter (baseline)
    ├── Cranelift (JIT, mobile-friendly)
    └── LLVM (future AOT)

Key Architectural Decisions

1. Dala Typed IR (Stable SSA)

The IR is the central hub. Every optimization, backend, and runtime feature flows through it:

• Stable tuple shapes: Fixed-layout tuples with known element types → compact native representation
• Message types: Expected message shapes for mailbox fast-path matching
• Actor types: Protocol-aware actor references with lifecycle metadata
• Tensor types: Shape + dtype for zero-copy AI interop
• Capability types: Typed native resource handles (GPU, files, sockets)

2. Semantic Layer vs Execution Backend

Clean separation between:

Semantic Layer (actors, mailboxes, supervisors, reductions, pattern matching, fault handling) Execution Backend (SSA lowering, register allocation, vectorization, machine code)

This gives huge long-term flexibility — swap the execution backend without changing actor semantics.

3. Specialized Mailbox System

Four priority queues (Critical, High, Normal, Low) with:

• Type-tag indexed fast-path for receive
• Stable message layouts for zero-copy delivery
• Back-pressure when queues are full

4. Multiple Memory Regions

Instead of a single BEAM-style heap:

Region	Purpose
Actor Heap	Short-lived BEAM terms, GC'd
Stable Immutable Region (SIR)	Long-lived, never rescanned (UI trees, configs, schemas)
Binary Region	Large binaries, refcounted
Tensor Region	GPU/NN buffers, zero-copy
Native Resource Region	Capability-tracked handles
Arena Allocators	Frame-scoped, bulk-free

5. QoS-Aware Scheduler

Designed for mobile AI workloads:

• Thermal-aware: Reduces inference priority when device is hot
• Battery-aware: Deprioritizes background work when battery is low
• QoS classes: Realtime, UserFacing, Utility, Background
• Inference-priority actors: Deadline-aware scheduling for AI workers

6. AI Runtime Layer

First-class runtime support for AI:

• Inference Workers: Dedicated workers with thermal throttling
• Tensor Resources: Managed GPU/ANE buffers with zero-copy interop
• Streaming Pipelines: Actor-driven real-time inference
• Model Lifecycle: Load, cache, version, unload with proper cleanup

7. Capability-Based Native Resources

Instead of arbitrary native handles:

• Actor-owned capabilities with reference tracking
• Supervised native resources
• Transferable ownership
• Automatic cleanup on actor termination

8. Pattern Matching Optimization

Typed pattern matching + AOT:

• Tagged dispatch for known message shapes
• Specialized mailbox matching
• Stable tuple destructuring without runtime type checks
• Branch merging for identical pattern arms

Crates

Crate	Description
dala_runtime	Core runtime: actors, scheduler, GC, mailboxes, memory regions, AI layer
dala_ir	Typed SSA IR with optimization passes
dala_beam_loader	Parses .beam files into IR
dala_codegen	Cranelift-based native code generation (JIT + AOT)
dala_dispatch	Module dispatch, hot code loading, export tables
dala_aot	CLI tool orchestrating the full pipeline

Building

# Build all crates
cargo build --release

# Build just the runtime
cd dala_runtime && cargo build --release

# Build with JIT support (default)
cargo build --features jit

# Build for AOT-only (iOS, no JIT)
cargo build --features aot

License

This project is licensed under the Mozilla Public License, Version 2.0 (MPL-2.0). See the LICENSE file for details.