Code Generation

Sounio supports multiple code generation backends so it can target:

fast local execution (JIT)
native artifacts (object files / shared libraries)
GPU kernels (PTX/SPIR-V/Metal)
debug information for diagnostics and tooling

Where It Lives

crates/souc/src/codegen/ — lowering and backend-specific generators
crates/souc/src/backend/ — runtime glue (native runtime, effect dispatch, GPU bridge)
crates/runtime/ — runtime support crate

Backends (Overview)

At the architecture level:

HLIR / MIR
  -> backend selection
    -> native artifacts (ELF/Mach-O/PE) or JIT
    -> GPU kernels (PTX/SPIR-V/MSL)
    -> debug info + source maps

Cranelift

Cranelift is used for fast compilation and JIT-style workflows.

Entry points commonly live under:

crates/souc/src/codegen/cranelift.rs
crates/souc/src/codegen/mir_cranelift.rs

Typical flow:

MIR (SSA)
  -> MIR optimizations (optional)
  -> Cranelift IR (CLIF)
  -> machine code (JIT) / object output (AOT, when supported)

Native (no LLVM)

There is a “native” backend path intended to avoid an LLVM dependency, emitting platform-specific artifacts (ELF/Mach-O/PE) through the compiler’s native runtime glue.

See:

crates/souc/src/backend/native/
crates/souc/src/backend/native/elf.rs (Linux) and platform-specific modules

LLVM (optional)

LLVM support is feature-gated and used for optimized and portable AOT compilation.

See:

crates/souc/src/codegen/llvm/

GPU

GPU code generation is implemented under:

crates/souc/src/codegen/gpu/

The compiler can lower “kernel-like” code into GPU IR forms. Actual execution depends on the runtime and available toolchains (e.g., CUDA toolkit).

Targets commonly referenced in the codebase:

Target	Notes
PTX	NVIDIA CUDA
SPIR-V	Vulkan/OpenCL-style ecosystems
MSL	Apple Metal shading language

Debug Info

Debug info and source mapping live under:

crates/souc/src/codegen/debug/

Scientific Features