Target Backends

BHC is designed for multiple targets. “Haskell” doesn’t have to mean “one runtime on one OS.”

Available Targets

Native

Server deployments and desktop applications.

bhc --target=native Main.hs
# or just:
bhc Main.hs

Status: Supported

• Full runtime features
• All profiles available
• Best performance for server workloads

WASI/WASM

Sandboxed compute and edge deployments.

bhc --target=wasi Main.hs

Status: Supported

• WebAssembly output
• WASI system interface
• Works with edge profile
• Suitable for serverless and browser environments

Limitations:

• Some runtime features restricted by WASI
• Edge profile recommended for size

GPU (CUDA/ROCm)

Accelerated compute for numeric workloads.

bhc --target=cuda Main.hs
bhc --target=rocm Main.hs

Status: Supported

• NVIDIA CUDA (PTX code generation)
• AMD ROCm (AMDGCN code generation)
• Automatic kernel fusion from Tensor IR
• Device memory management
• Works with numeric profile

RISC-V

Emerging architecture support.

bhc --target=riscv64 Main.hs

Status: Supported

• 32-bit and 64-bit RISC-V
• Native LLVM backend

Experimental

Additional backends under consideration.

Target Selection

Command line

# Native (default)
bhc Main.hs

# WASI/WASM
bhc --target=wasi Main.hs

# Specify output
bhc --target=wasi -o app.wasm Main.hs

Cross-compilation

BHC supports cross-compilation from your development machine to different targets.

# Compile WASM from macOS
bhc --target=wasi Main.hs -o app.wasm

# Run with wasmtime
wasmtime app.wasm

Target-Specific Considerations

Native Target

Full feature set:

• All runtime profiles
• Full FFI support
• Native threads and concurrency
• System calls unrestricted

WASI Target

Restricted environment:

• WASI system interface only
• No native threads (cooperative concurrency)
• File system access through WASI capabilities
• Network access through WASI sockets (when available)

Recommended setup:

bhc --target=wasi --profile=edge Main.hs

GPU Target

For numeric workloads with large data parallelism:

• Tensor IR operations automatically lowered to GPU kernels
• Device memory managed by runtime
• Host/device transfers optimized
• Kernel fusion reduces memory bandwidth

Recommended setup:

bhc --target=cuda --profile=numeric Main.hs

FFI Considerations

Foreign function calls differ by target:

For portable code, abstract FFI behind a platform layer.

Runtime Size

Different targets have different runtime sizes:

Sizes are approximate and depend on what features your code uses. GPU targets require platform-specific drivers (CUDA toolkit or ROCm).

Combining Targets and Profiles

The edge profile is designed for WASI and other constrained targets. GPU targets require the numeric profile for Tensor IR integration.