HieraSynth: A Parallel Framework for Complete Super-Optimization with Hierarchical Space Decomposition

Modern optimizing compilers generate efficient code but rarely achieve theoretical optimality, often necessitating manual fine-tuning. This is especially the case for processors with vector instructions, which can grow the instruction set by an order of magnitude. Super-optimizers can synthesize optimal code, but they face a fundamental scalability constraint: as the size of the instruction set increases, the length of the longest synthesizable program decreases rapidly. To help super-optimizers deal with large instruction sets, we introduce HieraSynth, a parallel framework for super-optimization that decomposes the problem by hierarchically partitioning the space of candidate programs, effectively decreasing the instruction set size. It also prunes search branches when the solver proves unrealizability, and explores independent subspaces in parallel, achieving near-linear speedup. HieraSynth is sufficiently efficient to run to completeness even on many hard problems, which means that it exhaustively explores the program space. This ensures that the synthesized program is optimal according to a cost model. We implement HieraSynth as a library and demonstrate its effectiveness with a RISC-V Vector super-optimizer capable of handling instruction sets with up to 700 instructions while synthesizing 7–8-instruction programs. This is a significant advancement over previous approaches that were limited to 1–3 instructions with similar instruction set sizes. Specifically, HieraSynth can handle instruction sets up to 10.66× larger for a given program size, or synthesize up to 4.75× larger programs for a fixed instruction set. Evaluations show that HieraSynth can synthesize code surpassing human-expert optimizations and significantly reduce synthesis time, making super-optimization more practical for modern vector architectures.