[Bug tree-optimization/84037] [8 Regression] Speed regression of polyhedron benchmark since r256643

[hubicka at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=84037

--- Comment #5 from Jan Hubicka <hubicka at gcc dot gnu.org> ---
I have added cost model for gather/scatter. Perhaps it only needs to be tuned
to be more expensive.
avx512 is not needed for gather, it is supported by haswell (avx2).