I am not totally convinced here, mostly because the alternatives are pretty load bearing right now: TOSA and Linalg aren’t really designed at their core to provide what arith did provide on tensor (linalg is a pretty big hammer when all you need is an add!).
On the other hand we’re getting closer to a state where arith is redundant with LLVM (I should follow up on [RFC] arithmetic vs llvm dialect …), and only multi-dimensional vectors would remain as a difference here.
So I agree that the current state is mixing two level of abstractions in the same dialect, and it’s pretty undesirable, but without a simple “tensor arith” dialect, I am a bit concerned about the impact on the ecosystem (do we expect users like Triton to take a dependency on linalg just to express basic arithmetic on tensors?).