LLM Translation of Compiler Intermediate Representation

Researchers introduce IRIS-14B, a 14-billion-parameter LLM fine-tuned to translate compiler intermediate representations between GCC's GIMPLE and LLVM IR, achieving up to 44 percentage points higher accuracy than existing state-of-the-art models. The approach demonstrates how LLMs can function as interoperability layers in hybrid compiler architectures, enabling cross-toolchain workflows without modifying existing compiler infrastructure.

The paper addresses a fundamental infrastructure challenge in software compilation: GCC and LLVM, the two dominant compiler ecosystems, use incompatible intermediate representations that prevent seamless cross-toolchain integration. Historically, engineers have relied on manual rule-based translators, which are costly to maintain and difficult to extend across programming languages. IRIS-14B represents a paradigm shift by applying machine learning to this problem, treating IR translation as a sequence-to-sequence task learned from paired examples extracted from real C code.

This work emerges from a broader trend of applying foundation models to structured technical domains. While LLMs have demonstrated capability in code generation and natural language processing, their application to deterministic compiler artifacts is less explored. The model's superior performance compared to models with parameters ranging from 13B to 1 trillion suggests that task-specific fine-tuning on domain data outweighs raw model scale—a finding with implications for enterprise AI deployment.

The practical impact centers on developer productivity and ecosystem interoperability. Developers could theoretically leverage GCC's sophisticated frontends with LLVM's optimization passes, or vice versa, without engineering custom bridges. This reduces barriers to language adoption and enables smaller compiler projects to benefit from established toolchain ecosystems. However, the approach introduces a new dependency: models must be maintained and retrained as compiler specifications evolve. The proposed hybrid neuro-symbolic architecture, where LLMs augment rather than replace traditional compiler passes, mitigates risks of non-deterministic behavior in production code generation.

Key challenges remain: handling edge cases, ensuring mathematical correctness of translated IRs, and managing model versioning across compiler updates. Success adoption depends on integration into standard workflows and community acceptance of LLM-assisted compilation infrastructure.

• →IRIS-14B outperforms much larger open-source LLMs by up to 44 percentage points on GIMPLE-to-LLVM IR translation tasks
• →The model enables cross-toolchain compiler workflows by functioning as an interoperability layer without modifying existing infrastructure
• →LLM-based IR translation represents a data-driven alternative to high-maintenance rule-based compiler translators
• →Hybrid neuro-symbolic compiler architectures can integrate LLMs for deterministic compilation while preserving traditional compiler guarantees
• →Task-specific fine-tuning demonstrates that domain expertise matters more than raw model scale for technical code transformation