[mlir] [mlir][linalg] Support subtracting accumulation in partial reduction … (PR #214033)

[Andrzej Warzyński]

banach-space wrote:

> I would honestly just copy-over the flags, create the issue, raise it, and tackle this once the discussion is settled, at least not implement it in the scope of this PR, to have a consistent state after this PR is merged. It hasn't bothered anyone till now, I doubt it will soon.

+1 Just sign-post that this is done for consistency with addition and that we should re-visit at a later time.

> The current tiling infrastructure currently just clones the operation, and the corresponding overflow flags with it. It does not really handle it.

You can verify using this example:
```mlir
func.func @generic_op_tensors(
  %arg0 : tensor<?x?x?xi32>, %arg1 : tensor<?x?x?xi32>) -> tensor<?x?x?xi32> {
  %c0 = arith.constant 0 : index
  %c1 = arith.constant 1 : index
  %c2 = arith.constant 2 : index
  %0 = tensor.dim %arg0, %c0 : tensor<?x?x?xi32>
  %1 = tensor.dim %arg0, %c1 : tensor<?x?x?xi32>
  %2 = tensor.dim %arg0, %c2 : tensor<?x?x?xi32>
  %3 = tensor.empty(%0, %1, %2) : tensor<?x?x?xi32>
  %4 = linalg.generic
    {indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d1, d2)>,
                      affine_map<(d0, d1, d2) -> (d0, d2, d1)>,
                      affine_map<(d0, d1, d2) -> (d2, d1, d0)>],
     iterator_types = ["parallel", "parallel", "parallel"]}
    ins(%arg0, %arg1 : tensor<?x?x?xi32>, tensor<?x?x?xi32>)
    outs(%3 : tensor<?x?x?xi32>) {
    ^bb0(%arg2 : i32, %arg3: i32, %arg4: i32):
      %5 = arith.addi %arg2, %arg3 overflow<nsw> : i32
      linalg.yield %5 : i32
    } -> tensor<?x?x?xi32>
  return %4 : tensor<?x?x?xi32>
}

module attributes {transform.with_named_sequence} {
  transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {
    %0 = transform.structured.match ops{["linalg.generic"]} in %arg1 : (!transform.any_op) -> !transform.any_op
    %1, %loops:3 = transform.structured.tile_using_for %0 tile_sizes [2, 3, 4] : (!transform.any_op) -> (!transform.any_op, !transform.any_op, !transform.any_op, !transform.any_op)
    transform.yield
  }
}
```

```bash
$ mlir-opt --transform-interpreter file.mlir
```

https://github.com/llvm/llvm-project/pull/214033