* Maxim Kuvyrkov:
This is a report from Linaro’s automatic benchmarking system, which tracks upstream components and automatically bisects performance regressions down to a single commit. We have recently enabled these emails to be sent to patch authors automatically.
It appears that after your patch SPEC CPU2006’s 410.bwaves slowed down by 4% on aarch64-linux-gnu when built with -O2 -flto. Any theories on what could’ve caused it?
Does the benchmark link against libpthread? (It doesn't matter if it is actually multi-threaded, only linking counts.)
If it does not link against libpthread, then the changes may have introduced additional atomic instructions. However, Adhemerval removed these atomic instructions again from the current development version. So if they were the cause, performance should be back to the baseline.
But I find it somewhat unlikely that the benchmark exercises this code.
This is one of the first reports, and we welcome feedback on improving reporting template.
Perhaps indicate more clearly that it's about a performance regression.
In the end, someone familiar with the benchmark and experience in performance work needs to look at differences in execution profiles, I expect.
Thanks, Florian