https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109971
--- Comment #10 from Kewen Lin <linkw at gcc dot gnu.org> --- (In reply to JuzheZhong from comment #9) > (In reply to Kewen Lin from comment #8) > > I did SPEC2017 int/fp evaluation on Power10 at Ofast and an extra explicit > > --param=vect-partial-vector-usage=2 (the default is 1 on Power), baseline > > r14-1241 vs. new r14-1242, the results showed that it can offer some > > speedups for 500.perlbench_r 1.12%, 525.x264_r 1.96%, 544.nab_r 1.91%, > > 549.fotonik3d_r 1.25%, but it degraded 510.parest_r by 5.01%. > > > > I just tested Juzhe's new proposed fix which makes the loop closing iv > > SCEV-ed, it can fix the degradation of 510.parest_r, also the miss > > optimization on cunroll (in #c5), the test failures are gone as well. One > > SPEC2017 re-evaluation with that fix is ongoing, I'd expect it won't degrade > > anything. > > Thanks so much. You mean you are trying this patch: > https://gcc.gnu.org/pipermail/gcc-patches/2023-May/620086.html ? Yes, it means that Richi's concern (niter analysis but all analyses relying on SCEV are pessimized) does affect the exposed degradation and failures. Thanks for looking into it. > > I believe it can improve even more for IBM's target. Hope so, I'll post the new SPEC2017 results once the run finishes. btw, the SPEC2017 run with --param=vect-partial-vector-usage=2 here is mainly to verify the expectation on the decrement IV change, the normal SPEC2017 runs still use --param=vect-partial-vector-usage=1 which isn't affected by this change and it beats the former in general as the cost for length setting up.
