From Intent to Silicon: Direct Neural Synthesis for Compiler-less Embedded Development

The traditional embedded software development lifecycle imposes a significant "Semantic-to-Silicon Gap." Developers must rely on a complex, rigid toolchain-comprising compilers, linkers, and generic startup routines-which introduces inevitable overhead and distances the programmer from the baremetal hardware. This letter introduces Direct Neural Synthesis (DNS), a paradigm shift that eliminates the compiler entirely. By treating the Instruction Set Architecture (ISA) as a native language, DNS leverages Large Language Models (LLMs) to translate structured Hardware Requirement Descriptions (HRD) directly into executable machine code. To ensure reliability, we integrate a Hardware-in-the-Loop (HIL) verification mechanism governed by a reinforcement learning loop. Experimental results on a RISC-V platform demonstrate that DNS achieves a 97.8% code size reduction for control-oriented tasks and an 84.8% reduction for algorithmic tasks compared to optimized GCC binaries. Crucially, we demonstrate that DNS enables "Zero-Overhead" execution by autonomously managing system resources-such as dynamic stack initialization-thereby obviating the need for generic runtime environments like crt0.o.