dcaballe wrote: We could also implement this lowering using a `vector.extract` + `vector.insert` for the time being. I would rather do that than opening the door to diverging lowerings or excluding ops for specific backends. https://github.com/llvm/llvm-project/pull/73915