RISC-V Compiler Performance:A Comparison between GCC and LLVM/clang

RISC-V is a new open-source instruction set architecture (ISA) that in De-cember 2016 manufactured its rst mass-produced processors. It focuses onboth eciency and performance and diers from other open-source architec-tures by not having a copyleft license permitting vendors to freely design,manufacture and sell RISC-V chips without any fees nor having to sharetheir modications on the reference implementations of the architecture.The goal of this thesis is to evaluate the performance of the GCC andLLVM/clang compilers support for the RISC-V target and their ability tooptimize for the architecture. The performance will be evaluated from ex-ecuting the CoreMark and Dhrystone benchmarks are both popular indus-try standard programs for evaluating performance on embedded processors.They will be run on both the GCC and LLVM/clang compilers on dierentoptimization levels and compared in performance per clock to the ARM archi-tecture which is mature yet rather similar to RISC-V. The compiler supportfor the RISC-V target is still in development and the focus of this thesis willbe the current performance dierences between the GCC and LLVM com-pilers on this architecture. The platform we will execute the benchmarks onwil be the Freedom E310 processor on the SiFive HiFive1 board for RISC-Vand a ARM Cortex-M4 processor by Freescale on the Teensy 3.6 board. TheFreedom E310 is almost identical to the reference Berkeley Rocket RISC-Vdesign and the ARM Coretex-M4 processor has a similar clock speed and isaimed at a similar target audience.The results presented that the -O2 and -O3 optimization levels on GCCfor RISC-V performed very well in comparison to our ARM reference. Onthe lower -O1 optimization level and -O0 which is no optimizations and -Oswhich is -O0 with optimizations for generating a smaller executable code sizeGCC performs much worse than ARM at 46% of the performance at -O1,8.2% at -Os and 9.3% at -O0 on the CoreMark benchmark with similar resultsin Dhrystone except on -O1 where it performed as well as ARM. When turn-ing o optimizations (-O0) GCC for RISC-V was 9.2% of the performanceon ARM in CoreMark and 11% in Dhrystone which was unexpected andneeds further investigation. LLVM/clang on the other hand crashed whentrying to compile our CoreMark benchmark and on Dhrystone the optimiza-tion options made a very minor impact on performance making it 6.0% theperformance of GCC on -O3 and 5.6% of the performance of ARM on -O3, soeven with optimizations it was still slower than GCC without optimizations.In conclusion the performance of RISC-V with the GCC compiler onthe higher optimization levels performs very well considering how young theRISC-V architecture is. It does seems like there could be room for improvement on the lower optimization levels however which in turn could also pos-sibly increase the performance of the higher optimization levels. With theLLVM/clang compiler on the other hand a lot of work needs to be done tomake it competetive in both performance and stability with the GCC com-piler and other architectures. Why the -O0 optimization is so considerablyslower on RISC-V than on ARM was also very unexpected and needs furtherinvestigation