Hi GCC devs, In LLVM we have been carrying out nightly runs of SPEC CPU 2017 and llvm-test-suite[1] for Clang on RISC-V hardware. We’ve been using it to find areas of improvement and to track performance regressions.
Recently we've also extended it to build and test GCC nightly, which is available to view here: https://cc-perf.igalia.com/db_default/v4/nts/machine/2. We’re using LNT[2] for this which will flag any changes in code size or execution time from the previous run. It also collects profiling data so you can click on a benchmark run and see where the cycles are spent alongside the disassembly. Some specifics: - The hardware tested is a Banana Pi BPI-F3 16GB, i.e. spacemit-x60 - The tested configuration is -march=rva22u64_v -O3 -flto - To reduce noise the tests are run 3 times and the minimum time is taken. Only one test is run at a time, pinned to a core with ASLR disabled. - In order to get the tests to finish running in a day, we are only testing the SPEC rate tests on the train datasets. We’ve found this to be a useful resource so we thought it was worth sharing over on this mailing list. We hope you will find it useful too. For the curious, it’s also possible to see the comparison between GCC and Clang here: https://cc-perf.igalia.com/db_default/v4/nts/3?compare_to=4. You are beating us in quite a few benchmarks, so we have some catching up to do ;) Luke Lau Compiler engineer at Igalia [1]: https://github.com/llvm/llvm-test-suite [2]: https://llvm.org/docs/lnt/
