https://gcc.gnu.org/bugzilla/show_bug.cgi?id=70046
--- Comment #3 from Richard Biener <rguenth at gcc dot gnu.org> ---
I've confirmed the regression to be caused by r230647
Estimated Estimated
Base Base Base Peak Peak Peak
Benchmarks Ref. Run Time Ratio Ref. Run Time Ratio
-------------- ------ --------- --------- ------ --------- ---------
410.bwaves 13590 180 75.7 * 13590 198 68.6 *
with BASE on r230646 and PEAK on r230647 using -Ofast -march=haswell on a
Intel(R) Core(TM) i5-4670T
I can even reproduce the difference w/o any -march thus with just -Ofast:
Estimated Estimated
Base Base Base Peak Peak Peak
Benchmarks Ref. Run Time Ratio Ref. Run Time Ratio
-------------- ------ --------- --------- ------ --------- ---------
410.bwaves 13590 176 77.1 * 13590 199 68.5 *
As expected the difference is in mat_times_vec_
Samples: 1M of event 'cycles', Event count (approx.): 1280690858409
39.22% bwaves_peak.amd bwaves_peak.amd64-m64-gcc42-nn [.] mat_times_vec_
33.60% bwaves_base.amd bwaves_base.amd64-m64-gcc42-nn [.] mat_times_vec_
IV differences are a mixed bag but number of IVs are different for the
nest, the slower case having much more IVs in the inner loop.
