A new LLM-based agent system can automatically convert standard C code into the restricted dialect that hardware synthesis tools actually accept.
High-level synthesis (HLS) tools — software that turns C code into chip hardware descriptions — accept only a narrow subset of C. Most C programs fail somewhere across four required pipeline stages: compilation, C simulation, synthesis, and co-simulation. Existing LLM tools for fixing this only cover the early stages and dump raw tool logs straight to the model, producing fixes that are fragile and hard to reproduce. The new system, tested on Xilinx Vitis, structures the problem as a closed loop: generate a fix, verify it, diagnose what broke, and repair it again until all four stages pass. The researchers also built a component called Progressive Mismatch Localization Chain, which narrows down exactly where simulation output diverges from expected results using log normalization and code tracing. A retrieval-augmented generation layer pulls from a self-updating pool of past repair patterns to inform each fix.
HLS is a meaningful bottleneck in hardware design: engineers write logic in C, but getting that code through a synthesis toolchain without manual rewriting is tedious and error-prone. A system that automates the conversion and verifies it end-to-end could shorten chip design cycles in real workflows. The authors report their system outperforms comparable state-of-the-art models, though independent benchmarks on production codebases remain to be seen.
The approach echoes a broader pattern in AI coding tools — wrapping LLMs in structured feedback loops rather than treating them as one-shot oracles — which is increasingly where the meaningful gains are coming from.