MatcoreDSL is a mathematical/HPC compiler project. Its experimental mdslc-region
compiler admits closed mathematical regions inside ordinary C++, preserving
immutable values, explicit host-storage effects and checked failure semantics.
It has a validated, bounded Linux x86-64 generated CPU execution path alongside
native and numerically legal provider choices.
Start with the compact current checkpoint, then the build/install and language guide and executable two-GEMM example. The foundation evidence distinguishes what was actually executed from inspection, compatibility and future work.
The new route is opt-in, uses coherent Clang/LLVM/MLIR 21.1.8, and has no frozen source API or ABI. Whole-region transformation/fusion, automatic storage reuse, GPU/NPU execution and BLAS parity are not established. Only the strict GEMM primitive currently traverses the generated MLIR/LLVM pipeline; the region orchestration is statically compiled from authenticated Matcore semantics.
The standalone compiler lives under compiler/. Existing mutating C++ GEMM,
CPU runtime/provider and Windows compatibility retain their separate contracts.
Root Python/JIT/native-extension code remains a separate Linux-tested compatibility surface;
context.md contains historical background, not the current compiler
checkpoint. Generated region execution is not claimed on Windows or accelerators.