On Wed, 12 May 2021 00:06:34 GMT, Sergey Bylokhov <s...@openjdk.org> wrote:
> Performance in one of the tests in the bimg_misc group is dropped by 20%(or > 5% of the group) after some unused code was removed from the libawt. I assume > the size of the lib became smaller and GCC heuristics were changed to do not > try to increase the size by inlining over some percent. I tested various > options but most of them affect different test cases in the opposite ways. > > While testing I have found that the most profitable option is to simply use > the "O3" optimization level, moreover, after offline discussion, we have > found that this optimization level was used before, but due to the typo, we > missed this option and fallback to the "LOW" level. > > Note this will change optimization level for all platforms in case of any > regressions we may try to implement this change for GGC only or we may try to > use just these options to solve this performance issue: " > -ftree-loop-vectorize -fversion-loops-for-strides " > > > Summary: > b12: > Number of tests: 13 > Overall average: 1222067.1646119351 > Best spread: 0.07% variance > Worst spread: 3.22% variance > (Basis for results comparison) > > o3: > Number of tests: 13 > Overall average: 1991716.540920269 > Best spread: 0.0% variance > Worst spread: 2.22% variance > Comparison to basis: > Best result: 314.15% of basis > Worst result: 97.23% of basis > Number of wins: 8 > Number of ties: 4 > Number of losses: 1 > > two options: > Number of tests: 13 > Overall average: 1971528.6542701898 > Best spread: 0.0% variance > Worst spread: 7.08% variance > Comparison to basis: > Best result: 312.05% of basis > Worst result: 77.27% of basis > Number of wins: 6 > Number of ties: 1 > Number of losses: 6 Marked as reviewed by erikj (Reviewer). ------------- PR: https://git.openjdk.java.net/jdk/pull/3990
