tblah wrote: > Hello! Did anyone notice a performance regression after this PR on exchange2? > I see it on both EPYC 9684X and grace. The slowdown is more than 1.5x. > > I verified that performance restores when adding `-fwrapv` with the latest > flang compiler. The compilation options are `-Ofast -march=native`. I will > look through LLVM dumps, but please let me know if it is a known problem.
My team doesn't closely monitor V2 performance. Another team looked into it. I'm told that initially there was a slowdown because function specialization was not performed on the last iteration of digits. Then that fixed after https://github.com/llvm/llvm-project/pull/118219 so we are happy on our end. However if you aren't using LTO then you won't be seeing the function specialization anyway. I'm told that the unspecialized version of digits was worse after this patch because of extra spills in the hot part of the code. To me that sounds similar to the bwaves regression the first time this patch landed. More info here: https://github.com/llvm/llvm-project/issues/117318 https://github.com/llvm/llvm-project/pull/110063 _______________________________________________ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
