[Mlir-commits] [mlir] [MLIR][Vector] Add Lowering for vector.step (PR #113655)

[Kunwar Grover]

Groverkss wrote:

> > Do we? Is there a use case for this? I just want to make that we are not moving code for the sake of ... moving code around 😅

We do actually have a usecase downstream. Sometimes, we see patterns like:

```
%a = vector.step : vector<128xindex>
%b = vector.extract_strided_slice %thread_id, %a : vector<4xindex>
```

If you know it's a vector.step, you can fold the step + slice into:

```
%a = vector.step : vector<4xindex>
%b = arith.constant ...
%c = arith.addi %b, %a
%d = arith.addi %c, %thread_id
```

So you don't have to materialize a big constant like vector<128xindex> into private memory. Having vector.step allows you to know how the arith.constant materializes.

https://github.com/llvm/llvm-project/pull/113655