Cross-ISA Execution Optimization in Virtualized Environments

Abstract

Supporting legacy software on emerging hardware platforms remains a persistent engineering challenge. While dynamic binary translation (DBT) systems provide functional compatibility, their performance degradation under full-system virtualization is often underestimated. In practice, translation latency, page-level synchronization, and indirect branch resolution contribute non-trivially to runtime overhead. In this work we examine these bottlenecks in a full Linux guest running heterogeneous workloads compiled for a non-native ISA. Instead of redesigning the entire translation pipeline, we introduce a staged optimization strategy: (i) trace-level speculative compilation for hot regions, (ii) translation cache locality enhancement using region clustering, and (iii) delayed exception synchronization to reduce redundant context switches. Implementation on an x86-64 host executing RISC-V binaries shows consistent reduction in translation stalls and branch misprediction penalties. Across SPEC CPU2017 integer workloads, we observe measurable improvements in sustained throughput without sacrificing architectural correctness. The study highlights that system-level DBT performance is governed as much by memory behavior as by translation heuristics.