Quantitative Methods for Continuously Improving RISC-V Compilers - Philipp Tomsich, VRULL

Quantitative Methods for Continuously Improving RISC-V Compilers - Philipp Tomsich, VRULL Up until now, most commercial implementations of RISC-V have focused on embedded use cases; with the emergence of RISC-V as an alternative to the ARM and Intel ecosystems, higher-end applications that include storage and the datacenter are moving into focus. This also shifts the perspective on target workloads and representative benchmarks. We show how quantitative methods can be used to assess the quality of code generation and identify and prioritize potential improvements based on hot-block analysis, dynamic instruction count metrics, and instruction histograms. The difference applicability to small benchmarks (such as Coremark, where this method identifies both a superoptimisation opportunity for the CRC functions and highlights the absence of conditional-move instructions in the RISC-V instructions set) and large benchmarks (such as SPEC, which helps to improve and mature the tools across the board) are demonstrate