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

Watch this 15-minute conference presentation from OOPSLA 2025 that introduces HieraSynth, a groundbreaking parallel framework designed to overcome the fundamental scalability limitations of super-optimizers when dealing with large instruction sets. Learn how researchers Sirui Lu from the University of Washington and Rastislav Bodík from Google DeepMind developed a novel approach that hierarchically partitions the space of candidate programs to effectively reduce instruction set complexity while maintaining theoretical optimality guarantees. Discover how HieraSynth addresses the critical challenge where increasing instruction set sizes dramatically reduce the length of synthesizable programs in traditional super-optimizers, particularly problematic for modern processors with extensive vector instruction capabilities. Explore the framework's innovative techniques including hierarchical space decomposition, intelligent pruning of unrealizable search branches through solver-based proofs, and parallel exploration of independent subspaces that achieves near-linear speedup. Understand how this approach enables complete super-optimization that exhaustively explores the program space to guarantee optimal results according to specified cost models. Examine the practical implementation as a library and its demonstration through a RISC-V Vector super-optimizer capable of handling instruction sets with up to 700 instructions while synthesizing 7-8 instruction programs, representing a significant advancement over previous approaches limited to 1-3 instructions with similar instruction set sizes. Analyze the performance improvements showing HieraSynth can handle instruction sets up to 10.66× larger for given program sizes or synthesize up to 4.75× larger programs for fixed instruction sets, with evaluation results demonstrating code synthesis that surpasses human-expert optimizations while significantly reducing synthesis time, making super-optimization more practical for modern vector architectures.