%26 = "tfl.slice"(%arg3, %22, %25) : (tensor<*xf32>, tensor<?xi32>, tensor<?xi32>) -> tensor<*xf32>
    %27 = "tfl.reshape"(%5, %cst_12) : (tensor<i32>, tensor<1xi32>) -> tensor<1xi32>
    %28 = "tfl.concatenation"(%27, %21) {axis = 0 : i32, fused_activation_function = "NONE"} : (tensor<1xi32>, tensor<?xi32>) -> tensor<?xi32>

  func.return %1 : tensor<1x1x257xi32>
  // CHECK: %[[CONST_DIM:.*]] = "tfl.pseudo_const"() <{value = dense<1> : tensor<1xi32>}> : () -> tensor<1xi32>
  // CHECK: %[[RED:.*]] = "tfl.sum"(%arg0, %[[CONST_DIM]]) <{keep_dims = true}> : (tensor<1x16384x257xi32>, tensor<1xi32>) -> tensor<1x1x257xi32>
}

// CHECK-LABEL: BatchmatmulToReduceSumF32
// CHECK-NOT: "tfl.batch_matmul"

  // CHECK-DAG: %[[IND_SLICE0_START:.*]] = "tf.Const"() <{value = dense<0> : tensor<1xi32>}> : () -> tensor<1xi32>
  // CHECK-DAG: %[[IND_SLICE0_SIZE:.*]] = "tf.Const"() <{value = dense<1> : tensor<1xi32>}> : () -> tensor<1xi32>
  // CHECK: %[[IND_SLICE0:.*]] = "tf.Slice"(%[[INDS]], %[[IND_SLICE0_START]], %[[IND_SLICE0_SIZE]]) : (tensor<2xi32>, tensor<1xi32>, tensor<1xi32>) -> tensor<1xi32>

    ```mlir
      %2 = "tf.A"(%arg0) : (tensor<?xi32>) -> tensor<?xi32>
      %3 = "tf.B"(%2) {device = "tpu0"} : (tensor<?xi32>) -> tensor<?xi32>
      %4 = "tf.C"(%2, %3) {device = "tpu0"} : (tensor<?xi32>, tensor<?xi32>) -> tensor<?xi32>
      %5 = "tf.D"(%4) : (tensor<?xi32>) -> tensor<?xi32>
    ```

    After the pass, we will have:

    ```mlir
      %0 = "tf.A"(%arg0) : (tensor<?xi32>) -> tensor<?xi32>

#loc3])\0Amodule {\0A  func.func @host_func(%arg0: tensor<3360x?xi32> loc(fused[#loc1, #loc2, #loc3]), %arg1: tensor<3xi32> loc(fused[#loc1, #loc2, #loc3]), %arg2: tensor<i32> loc(fused[#loc1, #loc2, #loc3]), %arg3: tensor<2xi32> loc(fused[#loc1, #loc2, #loc3])) -> (tensor<1x1120x?xi32>, tensor<1x1120x?xi32>, tensor<1120x?xi32>, tensor<2xi32>) {\0A    %0 = \22tf.Reshape\22(%arg0, %arg1) {_xla_outside_compilation = \220\22} : (tensor<3360x?xi32>, tensor<3xi32>) -> tensor<3x1120x?xi32> loc(#loc9)\0A    %1:3 = ...

# CHECK-NEXT:   "tfl.yield"(%[[OUTPUTS]]) : (tensor<*xi32>) -> ()
# CHECK-NEXT: }) {T = i32, device = ""} : (tensor<4xi32>, tensor<4xi32>) -> tensor<*xi32>
# CHECK-NEXT: return %[[CUSTOM]] : tensor<*xi32>

    %3 = "tf.Const"() {value = dense<2> : tensor<1xi32>} : () -> tensor<1xi32>
    // expected-remark@+1 {{Cost: 1}}
    %4 = "tf.Const"() {value = dense<1> : tensor<1xi32>} : () -> tensor<1xi32>
    // expected-remark@+1 {{Cost: 1}}
    %5 = "tf.Const"() {value = dense<0> : tensor<1xi32>} : () -> tensor<1xi32>
    // expected-remark@+1 {{Cost: 1}}
    %6 = "tf.Const"() {value = dense<3> : tensor<1xi32>} : () -> tensor<1xi32>

    %elem_1 = "tf.Const"() {value = dense<10.0> : tensor<8xf32>} : () -> tensor<8xf32>
    %tl_set_item = "tf.TensorListSetItem"(%tl, %idx, %elem_1) : (tensor<!tf_type.variant<tensor<*xf32>>>, tensor<i32>, tensor<8xf32>) -> tensor<!tf_type.variant<tensor<*xf32>>>
    %elem_shape_2 = "tf.Const"() {value = dense<-1> : tensor<1xi32>} : () -> tensor<1xi32>

// CHECK-DAG:   [[ZERO_1:%cst.*]] = arith.constant dense<0> : tensor<i32>
// CHECK-DAG:   [[PARTIAL_INDEX:%.*]] = "tf.Fill"([[VECTOR_RANK]], [[ZERO_1]]) : (tensor<1xi32>, tensor<i32>) -> tensor<?xi32>
// CHECK-DAG:   [[EXPANDED_INDIX:%.*]] = "tf.ExpandDims"([[INDEX]], [[ZERO]]) : (tensor<i32>, tensor<i32>) -> tensor<1xi32>

        %5:2 = "tf.C"(%4) : (tensor<2xi32>) -> (tensor<2xi32>, tensor<2xi32>)
        tf_device.return %5#0, %5#1 : tensor<2xi32>, tensor<2xi32>
      }) {num_cores_per_replica = 1, topology =  "", device_assignment =  []} : () -> (tensor<2xi32>, tensor<2xi32>)
      tf_device.return %2#1 : tensor<2xi32>
    }
    func.return %1 : tensor<2xi32>
  }