https://gcc.gnu.org/bugzilla/show_bug.cgi?id=119010
--- Comment #23 from Richard Biener <rguenth at gcc dot gnu.org> --- I'm not sure similar issues cause observed regression for icelake - bisecting that would help, it happened between r15-6757-gbd28244ec92b9a (good) and r15-6826-gc1729df6ec1eff (first known bad) when building with -Ofast -march=native - it might be because of the vectorizer early break improvements or the new hardreg PRE pass. My bet is on the vectorizer parts. r15-5918 size: text data bss dec hex filename 299920 0 476 300396 4956c ComputeNonbondedUtil.o r15-7037 size: text data bss dec hex filename 395480 0 476 395956 60ab4 ComputeNonbondedUtil.o and compile-time increases proportional to that. We now have 130 vectorized loops vs. 50 for the ComputeNonbondedUtil.C TU.
