Joint Performance and Energy Optimization for Cross-Architecture Binary Translation

Abstract

Dynamic binary translation remains a fundamental technique for enabling cross-architecture compatibility across heterogeneous computing platforms. However, conventional system-level translators rely on handcrafted translation rules and architecture-specific engineering, resulting in limited adaptability and substantial maintenance overhead when targeting emerging instruction set architectures.

This work introduces an adaptive cross-ISA execution framework that leverages machine-guided intermediate representation modeling to automate the generation of translation pipelines. Instead of manually designing architecture-dependent modules, the proposed system constructs an extensible translation graph from formal ISA specifications and applies hybrid static–dynamic optimization passes during runtime. The framework operates within a lightweight virtualization layer that allows aggressive just-in-time optimization while maintaining full-system compatibility. To further improve translation efficiency, we incorporate runtime workload profiling and adaptive code specialization mechanisms that dynamically refine hot execution paths. Unlike traditional DBT systems focused solely on throughput, our approach jointly optimizes execution latency, memory traffic, and energy-delay product (EDP). The framework is evaluated using SPEC CPU2017 and PARSEC benchmarks under a full Linux system environment. Experiments translating 64-bit RISC-V workloads to heterogeneous x86-based edge servers demonstrate an average performance improvement of 1.87× over mainstream DBT baselines, while reducing energy-delay product by 34%. For floating-point intensive applications, performance gains reach up to 4.12×. Beyond performance improvements, the proposed architecture significantly lowers the engineering complexity required to support new ISAs through automated specification-driven translation generation. The results indicate that adaptive, optimization-aware DBT frameworks can provide both scalability and efficiency for next-generation heterogeneous computing systems.