Large Language Model (LLM)-based end-to-end binary-to-C decompilation has advanced rapidly, yet current systems struggle to achieve robust data recovery, reliable global/static state reconstruction, and broad generalization. We present a modular framework that advances LLM decompilation along two axes. First, we introduce SynthDataBench, a programmatically generated C corpus comprising 1000 functions with 100 unit tests each, compiled with GCC and optimization levels O0–O3 to systematically evaluate global and static variable recovery. Second, we develop a semantics-aware binary extraction mechanism that jointly analyzes the .rodata and .data sections to recover values, types, and variable lifetimes with higher fidelity. This undertaking fundamentally depends on Supercomputing capabilities, involving the parallelized generation of synthetic corpora across distributed HPC clusters and the fine-tuning of Large Language Models on multi-GPU high-performance infrastructures to process the 0.4 billion token training corpus and execute 100,000 unit tests across multiple optimization levels. On SynthDataBench, our approach elevates re-executability from 17.5% to 86.2% (+ 68.7 percentage points) against ‘ReF Decompile’, a state-of-the-art model, while improving global/static variable recovery. Across the HumanEval benchmark under comparable compiler settings against ReF Decompile, our semantics-aware binary extraction-enhanced model obtains gains in compile rates and matches the run rates of the previous state-of-the-art model while supporting more features. These results establish a new state-of-the-art for data-aware LLM decompilation and demonstrate that comprehensive benchmark-driven training, enabled by large-scale parallel computing and combined with precise binary analysis, is essential for bridging the gap between syntactically plausible and semantically faithful decompilation.
