IRIS-14B: LLM-Based Translation of Compiler Intermediate Representations
Researchers have introduced IRIS-14B, a 14-billion-parameter transformer model designed to translate GCC's GIMPLE intermediate representation (IR) into LLVM IR. This innovation addresses significant interoperability barriers between major compiler infrastructures, GCC and LLVM, which traditionally rely on complex, high-maintenance rule-based translators. By leveraging a data-driven approach, IRIS-14B learns complex mappings directly from paired IRs extracted from C sources. Evaluations on real-world C code and competitive programming problems demonstrate that the model outperforms existing state-of-the-art open models, ranging from 13 to 1,000 billion parameters, by up to 44 percentage points in accuracy. As the first model explicitly trained for IR-to-IR translation, it enables cross-toolchain workflows without modifying existing compiler passes. This development supports hybrid neuro-symbolic compiler architectures, where LLMs act as interoperability layers while traditional infrastructure handles deterministic compilation. The findings suggest a promising future for integrating large language models into software infrastructure to enhance frontend and backend reuse across diverse programming ecosystems.
Wire timeline
IRIS-14B: LLM-Based Translation of Compiler Intermediate Representations
Researchers have introduced IRIS-14B, a 14-billion-parameter transformer model designed to translate GCC's GIMPLE intermediate representation (IR) into LLVM IR. This innovation addresses significant interoperability barriers between major compiler infrastructures, GCC and LLVM, which traditionally rely on complex, high-maintenance rule-based translators. By leveraging a data-driven approach, IRIS-14B learns complex mappings directly from paired IRs extracted from C sources. Evaluations on real-world C code and competitive programming problems demonstrate that the model outperforms existing state-of-the-art open models, ranging from 13 to 1,000 billion parameters, by up to 44 percentage points in accuracy. As the first model explicitly trained for IR-to-IR translation, it enables cross-toolchain workflows without modifying existing compiler passes. This development supports hybrid neuro-symbolic compiler architectures, where LLMs act as interoperability layers while traditional infrastructure handles deterministic compilation. The findings suggest a promising future for integrating large language models into software infrastructure to enhance frontend and backend reuse across diverse programming ecosystems.
cs.AI updates on arXiv.org