[Mlir-commits] [mlir] [mlir][ArmSME] Lower vector.extract/insert on SME tiles to MOVA intrinsics (PR #67786)

[Benjamin Maxwell]

================
@@ -496,3 +496,390 @@ func.func @vector_outerproduct_add_masked_f32(%lhs : vector<[4]xf32>, %rhs : vec
   %0 = vector.mask %mask { vector.outerproduct %lhs, %rhs, %acc {kind = #vector.kind<add>} : vector<[4]xf32>, vector<[4]xf32> } : vector<[4]x[4]xi1> -> vector<[4]x[4]xf32>
   "prevent.dce"(%0) : (vector<[4]x[4]xf32>) -> ()
 }
+
+//===----------------------------------------------------------------------===//
+// vector.insert
+//===----------------------------------------------------------------------===//
+
+// -----
+
+// CHECK-LABEL: @vector_insert_slice_i32(
+// CHECK-SAME:                       %[[TILE:.*]]: vector<[4]x[4]xi32>,
+// CHECK-SAME:                       %[[SLICE:.*]]: vector<[4]xi32>,
+// CHECK-SAME:                       %[[INDEX:.*]]: index)
+func.func @vector_insert_slice_i32(%tile: vector<[4]x[4]xi32>, %slice: vector<[4]xi32>, %row: index) -> vector<[4]x[4]xi32>{
+  // CHECK-NEXT: %[[PTRUE:.*]] = arith.constant dense<true> : vector<[4]xi1>
+  // CHECK-NEXT: %[[TILE_ID:.*]] = arm_sme.cast_vector_to_tile %[[TILE]] : vector<[4]x[4]xi32> to i32
+  // CHECK-NEXT: %[[TILE_SLICE_INDEX:.*]] = arith.index_castui %[[INDEX]] : index to i32
+  // CHECK-NEXT: "arm_sme.intr.write.horiz"(%[[TILE_ID]], %[[TILE_SLICE_INDEX]], %[[PTRUE]], %[[SLICE]]) : (i32, i32, vector<[4]xi1>, vector<[4]xi32>) -> ()
+  %new_tile = vector.insert %slice, %tile[%row] : vector<[4]xi32> into vector<[4]x[4]xi32>
+  return %new_tile : vector<[4]x[4]xi32>
+}
+
+// -----
+
+// CHECK-LABEL: @vector_insert_slice_i128
+func.func @vector_insert_slice_i128(%tile: vector<[1]x[1]xi128>, %slice: vector<[1]xi128>, %row: index) -> vector<[1]x[1]xi128> {
----------------
MacDue wrote:

I've kept `i32` as the first in each group as it's the only full test. The rest are just for completeness. 

https://github.com/llvm/llvm-project/pull/67786