Magpie

This experimental programming language and compiler toolchain focuses on deterministic source and machine output, explicit ownership and borrowing rules with ARC‑managed heap lifetimes, and strong diagnostics for rapid error resolution. It offers a command‑line interface designed for automation and large‑language‑model agents, enabling fast, repeatable build workflows. The compiler pipeline includes lexer, parser, semantic analysis, type checking, ownership checking, mid‑level IR lowering, monomorphization, and ARC insertion passes. It can generate code for CPU backends (LLVM, WASM) and multiple GPU backends (SPIR‑V, MSL, PTX, HIP, WGSL), with runtime support for dynamic dispatch across Metal, CUDA, HIP, Vulkan, and WebGPU. Additional features include a bfloat16 primitive, MLX host API integration for Apple Silicon acceleration, and a GPU profiling system that exports Chrome trace data and tracks allocations. The toolchain is aimed at systems programmers, language researchers, and GPU developers who need reproducible builds, efficient GPU kernels, and seamless ML integration.