libloong

LoongArch64 sandboxing library

libloong is a 64-bit high-performance LoongArch emulator library designed for embedding and high-frequency scripting. Built on the architecture of libriscv, it provides the fastest 64-bit interpreter performance available, maintaining a compact ~18k LOC codebase with zero dependencies.

You can also run LoongArch64 Linux programs in the command-line interface.

• Universal: Unlike JITs, which require platform-specific code generation and are unavailable on locked-down platforms, libloong runs anywhere. This includes Nintendo Switch (1 & 2), iOS, and other environments.
• Standard: Runs standard Linux binaries.
• Ultra-Low Latency: Traditional VMs like Lua or Java often suffer from high call overhead (~150ns). libloong achieves a ~4ns native-to-guest call overhead, making it suitable for hot-loop game engine scripting.

Performance Benchmarks

In CoreMark 1.0 interpreter benchmarks (December 2025), libloong leads the 64-bit category on the Ryzen 7950X, reliably reaching a score of 3000+.

CoreMark 1.0 Interpreter Scores

Interpreter	Architecture	Score
libloong	64-bit LoongArch	3045
libriscv	64-bit RISC-V	2865
stitch	Wasm	2743
wasm3	Wasm	2368
wasmer (WAMR)	Wasm	2314
wasmi	Wasm	1967

Advanced Modes: While the interpreter is the primary focus for portability, libloong includes a lightweight JIT reaching 38% of native (15.5k CoreMark) and an embedded binary translator reaching ~77% of native (31.9k CoreMark).

Features

• Bindings: Native C++ API with Rust and Go bindings.
• Vector Support: Full support for LSX and LASX instruction sets.
• Memory Safety: Strict memory sandboxing with optional masked memory arenas.
• State Management: First-class support for pausing, resuming, and serializing machine state.
• Dynamic execution: Can load and execute dynamic executables, as well as embedded JITs (eg. LuaJIT)*

Design

Game engine scripting is where libloong excels. Traditional games expose modding through shared libraries (full system access), embedded VMs like Lua (~150ns call overhead), or Java run-times. libloong has ~4ns call overhead.

See the example Asteroid game.

Building

CMake configuration options:

• LA_DEBUG=ON/OFF - Enable debug output (default: OFF)
• LA_BINARY_TRANSLATION=ON/OFF - Enable binary translation (default: OFF)
• LA_THREADED=ON/OFF - Enable threaded bytecode dispatch (default: ON)
• LA_MASKED_MEMORY_BITS=N - Set masked memory arena size to 2^N bytes (0 = disabled, default: 0)

Example with options:

cmake .. -DCMAKE_BUILD_TYPE=Release \
         -DLA_MASKED_MEMORY_BITS=32 \
         -DLA_BINARY_TRANSLATION=ON
make -j6

Quick Start

#include <libloong/machine.hpp>

int main() {
    // Load a LoongArch ELF binary
    std::vector<uint8_t> binary = load_file("program.elf");

    // Create a machine with 64MB memory
    loongarch::Machine machine { binary, {
        .memory_max = 64 * 1024 * 1024
    }};

    // Setup program arguments
    machine.setup_linux({"program"}, {"LC_ALL=C"});

    // Run the program
    machine.simulate();
}

Documentation

• Integration Guide
• API Reference
• LoongArch ISA Support