[Mlir-commits] [mlir] [mlir][ArmSME] Refactor ArmSMEToSCF to used shared loop-building helper (NFC) (PR #79172)

[Andrzej Warzyński]

================
@@ -47,99 +48,92 @@ void getMemrefIndices(ValueRange indices, unsigned rank, Value tileSliceIndex,
 
   if (rank == 2)
     outIndices.push_back(indices[1]);
-}
 
-/// Lower `arm_sme.tile_load` to a loop over the tile slices and load each slice
-/// using `arm_sme.load_tile_slice`.
-///
-///  BEFORE:
-///  ```mlir
-///  %tile = arm_sme.tile_load %src[%c0, %c0] :
-///    memref<?x?xi32>, vector<[4]x[4]xi32>
-///  ```
-///
-///  AFTER:
-///  ```mlir
-///  %ptrue_s = arith.constant dense<true> : vector<[4]xi1>
-///  %init_tile = arm_sme.get_tile : vector<[4]x[4]xi32>
-///  %vscale = vector.vscale
-///  %c0 = arith.constant 0 : index
-///  %c1 = arith.constant 1 : index
-///  %min_svl_s = arith.constant 4 : index
-///  %svl_s = arith.muli %min_svl_s, %vscale : index
-///  %tile = scf.for %tile_slice_idx = %c0 to %svl_s step %c1
-///                iter_args(%iter_tile = %init_tile) -> (vector<[4]x[4]xi32>) {
-///    %tile_update = arm_sme.load_tile_slice %src[%tile_slice_idx],
-///      %ptrue_s, %iter_tile, %tile_slice_idx
-///        : memref<?x?xi32>, vector<[4]xi1>, vector<[4]x[4]xi32>
-///    scf.yield %tile_update : vector<[4]x[4]xi32>
-///  }
-///  ```
----------------
banach-space wrote:

I would love for this comment to be preserved in some form. I feel that we focus too much on finer details. Something like this would be, IMHO, sufficient:
```
///  %vscale = vector.vscale
///  %min_svl_s = arith.constant 4 : index
///  %svl_s = arith.muli %min_svl_s, %vscale : index
///  %tile = scf.for %tile_slice_idx = 0 to %svl_s step 1 iter_args(%iter_tile = %init_tile) -> (vector<[4]x[4]xi32>) {
///    %tile_update = arm_sme.load_tile_slice %src[%tile_slice_idx]
///    scf.yield %tile_update
///  }
```

As in, the overall structure is super helpful. The code itself can help with the finer details (like types).

https://github.com/llvm/llvm-project/pull/79172