llvmnet

Clang/Flang 22 -> LLVM IR -> unsafe .NET 10 CIL.

This is a working experimental compiler, not an emscripten-equivalent finished toolchain. C, C++, and Fortran share a direct CIL backend. Generated programs execute translated methods on .NET; they do not invoke native ffmpeg, interpret LLVM IR, or embed a WebAssembly engine.

The initial focus is feature correctness and single-process execution. The supplied software ffmpeg profile builds with dav1d, libxml2, and zlib. It has decoded all 1,150 frames of the supplied 4K HDR HEVC video with the same active-pixel hashes as native clang. Representative SPEC development workloads run after that ffmpeg gate. See docs/status.md for precise coverage and limitations.

A relocatable clang/flang-style distribution, two explicit ABI modes, NativeAOT output, and ordinary SPEC harness configuration are now implemented. See docs/toolchain.md for contracts, deployment requirements, and current limits.

Build

Requirements for the current Linux toolchain:

• .NET 10 SDK and runtime.
• LLVM/Clang/Flang 22, development headers, libLLVM-22.so, llvm-config-22, opt-22, llvm-ar-22, and llvm-ranlib-22.
• Bash and a C++ compiler (clang++-22) for the compile-time LLVM bridge.
• NuGet access for pinned MathNet.Numerics 5.0.0, used for erf.
• For dependency integration: CMake, Ninja, Meson, pkg-config, and initialized source checkouts.

dotnet build -c Release
bin/llvmnet tests/c/runtime.c -o artifacts/example.dll
dotnet artifacts/example.dll payload

Create the output directory first if it does not exist. Keep the emitted assembly, its runtime configuration, and the copied runtime/dependency assemblies together. LLVM and the compiler's native bridge are needed at compile time, not to run the result.

Compiler Modes

bin/llvmnet -O1 -c source.c -o source.o
llvm-ar-22 rcs libexample.a source.o
bin/llvmnet main.c -L. -lexample -o program.dll
dotnet program.dll argument

bin/llvmnet -std=c++17 program.cpp -o cpp-program.dll
bin/llvmnet program.f90 -o fortran-program.dll
bin/llvmnet -S -emit-llvm source.c -o source.ll
bin/llvmnet main.c libexample.a --emit-llvm -o program.bc

Objects and archives contain LLVM bitcode, not native object code. Explicit object files are linked eagerly; archive members are selected to satisfy unresolved symbols, with rescanning for inter-archive dependencies. -lc and -lm select managed exports in managed modes and real system libraries in system mode.

--keep-ir linked.bc preserves linked IR for investigation. LLVMNET_TRACE=1 enables compiler exception traces. LLVMNET_CLANG and LLVMNET_FLANG override frontend executables. Use bin/llvmnet --help for the supported driver interface; it is not yet a complete clang/emcc flag replacement.

Managed Versus Native Linking

These are deliberately different operations:

bin/llvmnet --emit-library library.c -o Example.dll
bin/llvmnet main.c --cil-reference Example.dll -o managed-client.dll

bin/llvmnet -fno-builtin tests/c/native-import.c \
  --native-import 'abs=libc.so.6!abs' -o native-client.dll

Managed references resolve public static methods annotated with LlvmNet.Runtime.CExportAttribute, also emitted on translated methods. Scalar signatures and native-width pointer values must match. Cross-assembly aggregate signatures are not yet a stable public ABI.

The explicit --native-import path creates opt-in P/Invoke declarations, one symbol mapping per argument. Only non-variadic scalar C signatures are supported by this path; aggregate/by-value/sret signatures are rejected. This path does not supply native callbacks or native data imports. There is no implicit native-library fallback in the managed modes.

That restriction applies to the explicit scalar --native-import path. The newer --runtime=system mode uses LLVM-generated native ABI bridges and supports tested C struct/varargs calls, native data imports, and scalar callbacks. Native C++ library ABI interchange remains unsupported.

Deploy and Select an ABI

bash scripts/package-toolchain.sh /path/toolchain --self-contained --bundle-llvm
/path/toolchain/bin/llvmnet-clang --runtime=system source.c --nativeaot -o program
/path/toolchain/bin/llvmnet-clang --runtime=portable source.c -o portable.dll
/path/toolchain/bin/llvmnet-clang++ --runtime=portable source.cpp -o cpp.dll
/path/toolchain/bin/llvmnet-flang --runtime=portable source.f90 -o fortran.dll

system uses the native Linux x86-64 ABI and libc/pthread interfaces. portable uses the bundled, versioned dotnet64-v1 ABI and managed adapters, with libc++ headers/support bitcode included. The existing managed-host default remains a legacy compatibility mode, not a cross-platform ABI. Objects and managed references from different modes are rejected.

Outputs without .dll are ordinary Unix executable launchers plus a payload directory; --nativeaot produces a native executable with .NET runtime code included. NativeAOT publishing needs the .NET 10 SDK and native linker even when the compiler installation is self-contained. Cross-platform execution of portable CIL is a release goal with a fixed ABI, but has only been tested on Linux x86-64 so far.

Tests and Workloads

bash tests/all.sh core
bash tests/all.sh integration
bash tests/toolchain-all.sh

bash scripts/build-ffmpeg.sh
bash tests/codecs.sh
bash tests/codecs-extended.sh
bash tests/manifests.sh
FFMPEG_THREADS_AOT=1 bash tests/ffmpeg-threads.sh
bash tests/wasm-simd128.sh
bash scripts/benchmark-video.sh

bash scripts/spec-mcf.sh test
bash scripts/spec-mcf.sh refrate
bash scripts/spec-deepsjeng.sh test
bash scripts/spec-exchange2.sh test
bash scripts/spec-zstd.sh
bash scripts/spec-lbm.sh 2017 test
bash scripts/spec-lbm.sh 2026 test

Run scripts/build-dependencies.sh before the full ffmpeg build. Its defaults use the supplied LightPlayer-Packages checkout and cached zlib 1.3.2. Override LIGHTPLAYER_PACKAGES, FFMPEG_SOURCE, LLVMNET_DEPS_PREFIX, or ZLIB_SOURCE when needed. The smaller scripts/build-ffmpeg-core.sh is useful for backend bring-up.

The full ffmpeg profile enables pthreads. Select LLVMNET_RUNTIME=system or LLVMNET_RUNTIME=portable for both dependency and ffmpeg builds; their default outputs are isolated by ABI. Add FFMPEG_SIMD128=1 to the ffmpeg build for HEVC SIMD128 through .NET vectors, with separate -simd128 outputs. The decoder accepts input [frame-limit] [video|audio] [threads:1-64] [auto|frame|slice] [hash|bench]; bench skips hashing. See docs/ffmpeg-threading.md, docs/simd128.md, and the six-video before/after measurements. JIT SIMD improves throughput; baseline NativeAOT SIMD currently regresses, while explicit host-targeted AOT improves it.

Build outputs, generated media, and locally extracted SPEC materials stay under ignored artifacts/. The scripts do not modify the original source checkouts or mounted SPEC kits. SPEC development scripts stage only the relevant licensed benchmark archives. A standard harness runner and configuration are also available via scripts/spec-harness.sh. These checks are not full-suite SPEC-reportable runs or scores.

Safety and Portability

Generated code is intentionally unsafe and is not a sandbox. C pointers, invalid memory accesses, or untrusted programs can corrupt memory or terminate the process. Only compile and execute code you trust.

Host-ABI modes target Linux/x86-64 SysV. Portable mode uses a fixed little-endian 64-bit .NET ABI with owned headers and runtime structures, not host headers. None of the modes is a sandbox or a complete libc/POSIX implementation. See docs/toolchain.md, docs/architecture.md, and docs/roadmap.md.