https://gcc.gnu.org/bugzilla/show_bug.cgi?id=71361
amker at gcc dot gnu.org changed: What |Removed |Added ---------------------------------------------------------------------------- CC| |amker at gcc dot gnu.org --- Comment #1 from amker at gcc dot gnu.org --- It doesn't look like a regression from the dump, I suspect it's because how gcc handles symbol (arr_1/arr_2) in m32 PIE code. I will have a look. BTW, the patch itself is right, it triggers cost model issue again in which wrong/inaccurate cost gives better result. I am doing experiments rewriting the whole cost computation part.