There is a trunc(lshr (zext A), Cst) optimization in InstCombineCasts that