AI summary by TwinMind
JS At The Speed Of C
Theo - t3․gg
27m 54s
JavaScript Performance Limitations
- JavaScript isn't fastest language by quite a bit
- Performance issues in:
- Processing lots of requests (non-serverless)
- Video encoding
- Game development
- Complex animations
- Crypto hashing/Bitcoin mining
Static Hermes Overview
- Next version of Hermes JavaScript engine developed by Meta
- Code-named "static Hermes" - open on GitHub since day one
- Main innovation: optional ahead-of-time compilation of JavaScript to native code
- Goal: combine predictability/high performance of C with JavaScript usability
- Partnership with Amazon for TypeScript support (Meta uses Flow internally)
Technical Architecture
- Not for entire app compilation - selective optimization
- Choose specific parts needing maximum performance
- Rest continues using Hermes bytecode
- Relies on sound types for native performance
- Modifies some JavaScript semantics to enable efficient sound typing
- Runtime types guaranteed to match type annotations
- Completely safe by default - no native crashes
Performance Benchmarks
- nbody JS benchmark results:
- Initial: 550ms (same as Hermes)
- Static Hermes: 10x faster performance
- Potential for 20x faster than Hermes
- Before: dozens of instructions per loop line via interpreter
- After: 6 assembler instructions for first two loop lines
- Most calculations performed in parallel by single instruction
- Performance close to C code
Demo Implementation
- JavaScript code calling C++ game library (Dear ImGui) and 3D backend (Sokol)
- Imperative game UI library rendering meshes
- 100% compiled native code JavaScript
- Entire UI state recreated every frame
- Performance comparison:
- 120 Hz on MacBook
- 60 Hz on old Motorola G5 phone
- C++ and JavaScript versions run with same performance
- ~1000 lines of JavaScript code with static Hermes built-ins
Current Status & Limitations
- Work in progress - not suitable for general use
- Only for extreme hackers/enthusiasts currently
- Not easy to use, many things broken
- Encouraging results but still experimental
- Team actively working on improvements
- Community feedback desired
Technical Implementation Details
- Uses type information to generate efficient native code
- Dynamic code (number or string) vs static optimizations
- Sound types prevent runtime type mismatches
- Foreign Function Interface (FFI) for native library access
- Portable across all platforms
- ARM64 assembler output with minimal instructions
Comparison with Other Solutions
- Similar to Go's Arenas feature for specific memory management
- Selective performance optimization vs rewriting everything
- Flash List library example for React Native performance improvements
- Shopify's shop.app performance achievements using React Native
