On Fri, Oct 25, 2024 at 12:22 AM Kugan Vivekanandarajah <kvivekana...@nvidia.com> wrote: > > Hi, > > This patch sets param_vect_max_version_for_alias_checks to 15. > This was causing GCC to miss vectorization opportunities in one internal > application making it slower than LLVM by about ~14%. > > I've tested different param_vect_max_version_for_alias_checks such as 15 and > 100 > and the SPEC2017 results are neutral.
Can you share code size numbers as well? > I have also tested with some other > thirdparty benchmarks and the results are similar. > > Original default of 10 itself seem arbitary. Given that, GCC's vectoriser does > consideres cost of alias checks, increasing this param seems reasonable. I'll note the "cost" is quite rough at the moment but we are factoring this into the minimum number of scalar iterations test and we make sure to test that first with a branch (we didn't in the past), so I agree. > In our case we need a value of at teast 11 where as the current > default is 10. > > Bootstrapped and regression tested on aarc64-linux-gnu with no new > regressions. OK. Thanks, Richard. > Thanks, > Kugan > > >
