https://gcc.gnu.org/bugzilla/show_bug.cgi?id=127205
--- Comment #5 from Pengfei Li <pfustc at gcc dot gnu.org> --- Let me correct what I said above. After the patches, there are more IVs in the outer loop, but the innermost loop ends up using fewer IVs - just a shared offset. This is less profitable than before. I don't fully understand why the cost computation changes lead to less IVs in the innermost loop.
