[Mlir-commits] [mlir] [MLIR][AArch64] Add an extra test for Neon I8MM (NFC) (PR #135777)

[Andrzej Warzyński]

================

----------------
banach-space wrote:

Thanks for the update! I still find this test quite difficult to follow - there’s just too much going on for it to be easily useful in its current form. It would benefit from either trimming or adding more structure/annotation. I was hoping for a more extensive refactor to improve readability.

Here are a few concrete suggestions:

* Use descriptive LIT variables for function arguments: `%arg0: vector<4x16xi8>` --> `%[[LHS:[0-9]+]]: vector<4x16xi8>` .
* Separate and annotate each iteration clearly - a blank line between iterations would help visually.
* Consider trimming verbose lines (example below).

BEFORE:
```mlir
// CHECK-NEXT: %[[VAL_48:[0-9]+]] = vector.extract_strided_slice %arg0 {offsets = [2, 0], sizes = [2, 8], strides = [1, 1]} : vector<4x16xi8> to vector<2x8xi8>
```
AFTER:
```
// CHECK-NEXT: %[[LHS_V2:[0-9]+]] = vector.extract_strided_slice [[LHS]] {offsets = [2, 0] {{.*}} to vector<2x8xi8>
```

To me, the key thing to validate here is that the correct **offsets** are used - that's the core behaviour this test is trying to verify. I'd also suggest removing the `shape_cast` ops unless they are absolutely necessary for this specific test, since their correctness is covered elsewhere.

https://github.com/llvm/llvm-project/pull/135777