[Mlir-commits] [mlir] [mlir][TilingInterface] Use `LoopLikeOpInterface` in tiling using SCF to unify tiling with `scf.for` and `scf.forall`. (PR #77874)

[Mehdi Amini]

================
@@ -218,6 +221,53 @@ def LoopLikeOpInterface : OpInterface<"LoopLikeOpInterface"> {
         return ::mlir::failure();
       }]
     >,
+    InterfaceMethod<[{
+        Append the specified additional "init" operands: replace this loop with
+        a new loop that has the additional init operands. The loop body of
+        this loop is moved over to the new loop.
+
+        This method is similar to `replaceWithAdditionalYields` but instead of
+        returning the value that is actually yielded, this returns the tiles of
+        the values that are yielded. This allows for unified handling of opreations
+        like `scf.forall` which dont yield a value from the loop, but instead
+        the terminator specifies where to insert the tile yielded by the body of
+        the loop. For example,
+        
+        ```mlir
+        %0 = scf.forall ... shared_outs(%arg0 = %arg1) {
+          ...
+          %tiled_value
+          scf.forall.in_parallel {
+            tensor.parallel_insert_slice %tiled_value into %arg0[%o1, %o2]...
+          }
+        }
+        ```
+
+        For an `scf.for` the same computation would be represented as
+        ```mlir
+        %0 = scf.for ... iter_args(%arg0 = %arg1) {
+          ...
+          %tiled_value
+          %insert = tensor.insert_slice %tiled_value into %arg0[%o1, %o2]...
+          scf.yield %insert
+        }
+        ```
+
+        So for the caller, the tiled value (`%tiled_values`) and the offsets
----------------
joker-eph wrote:

```suggestion
        So for the caller, the tiled value (`%tiled_value`) and the offsets
```

https://github.com/llvm/llvm-project/pull/77874