[llvm-branch-commits] [mlir] [MLIR] Integration tests for lowering vector.contract to SVE FEAT_I8MM (PR #140573)

[Andrzej Warzyński via llvm-branch-commits]

banach-space wrote:

Thanks - great to finally be reaching this stage! I have a few high-level questions and suggestions:

**1. Why is the scalable dimension always [4]?**

>From the current tests, it looks like the scalable dim is always `[4]`. Could you remind me why that value is chosen?

**2. Reduce duplication in the 4x8x4 tests**

The current tests differ only in terms of **input**/**output** and `extsi` vs `extui`. It should be possible to reduce duplication by extracting shared logic into helpers, and writing 4 separate entry points (set via `entry_point`) to isolate the differences.

For example:
```mlir
func.func @main_smmla() {
  // Init LHS, RHS, ACC
  // CHECK-LINES for LHS
  print(lhs);
  // CHECK-LINES for RHS
  print(rhs);
 
  arith.extsi (lhs)
  arith.extsi (rhs)
  vector.contract
  
  // CHECK-LINES for ACC
  print(acc);
}
```
This would keep the test logic focused and easier to maintain.

**3. Add checks for generated IR (LLVM dialect)**

It would be good to verify that the lowered IR includes the correct SME MMLA intrinsics. For example:

```mlir
// CHECK-COUNT-4: llvm.intr.smmla
```
This would help confirm both correctness and that the expected number of operations are emitted.

**4. Consider toggling VL within tests**
Have you considered toggling the scalable vector length (`VL`) within the test? That would allow verifying behaviour for multiple `VL` values.

>From what I can tell, this would only work if the inputs are generated inside a loop, similar to this example: https://github.com/llvm/llvm-project/blob/88f61f2c5c0ad9dad9c8df2fb86352629e7572c1/mlir/test/Integration/Dialect/Vector/CPU/ArmSME/load-vertical.mlir#L19-L37

That might be a nice validation of the "scalability" aspect.

https://github.com/llvm/llvm-project/pull/140573