%split_dim_2 = arith.constant dense<1> : tensor<1xi32>
  %4, %5 = "tfl.split"(%split_dim_2, %arg0) {num_splits = 2 : i32} : (tensor<1xi32>, tensor<16x4xf32>) -> (tensor<16x2xf32>, tensor<16x2xf32>)
  %6:2 = "tfl.split"(%split_dim_2, %arg0) {num_splits = 2 : i32} : (tensor<1xi32>, tensor<16x4xf32>) -> (tensor<16x2xf32>, tensor<16x?xf32>)

  %0 = "tf.Mean"(%arg0, %arg1) : (tensor<2x2xf32>, tensor<1xi32>) -> tensor<1x2xf32>
  func.return %0 : tensor<1x2xf32>

// CHECK-LABEL: mean
// CHECK:  "tfl.mean"(%arg0, %arg1) <{keep_dims = false}> : (tensor<2x2xf32>, tensor<1xi32>) -> tensor<1x2xf32>
}

func.func @mean_true(%arg0: tensor<2x2xf32>, %arg1: tensor<1xi32>) -> tensor<1x2xf32> {

    %0 = "tf.BatchMatMulV2"(%arg3, %arg1) {adj_x = false, adj_y = false} : (tensor<1x3xf32>, tensor<3x4xf32>) -> tensor<1x4xf32>
    %1 = arith.constant dense<[[2.3, 3.4, 4.5, 5.5]]> : tensor<1x4xf32>
    %2 = "tf.Add"(%0, %1) : (tensor<1x4xf32>, tensor<1x4xf32>) -> tensor<1x4xf32>
    %3 = tensor.cast %2 : tensor<1x4xf32> to tensor<1x?xf32>
    func.return %3 : tensor<1x?xf32>
}

// CHECK-LABEL: testFoldEnsureShapeOp
func.func @testFoldEnsureShapeOp(%arg0: tensor<10x20xf32>) -> (tensor<10x20xf32>, tensor<10x20xf32>, tensor<20x10xf32>) {
  %0 = "tf.EnsureShape"(%arg0) {shape = #tf_type.shape<10x20>} : (tensor<10x20xf32>) -> tensor<10x20xf32>
  %1 = "tf.EnsureShape"(%arg0) {shape = #tf_type.shape<?x20>} : (tensor<10x20xf32>) -> tensor<10x20xf32>
  // Failing case which should not be folded.

// Tests that a float `stablehlo.concatenate` is not converted to
// `tfl.concatenation`.

func.func @concatenate_float(%arg0: tensor<3x2xf32>, %arg1: tensor<1x2xf32>) -> tensor<4x2xf32> {
  %0 = "stablehlo.concatenate"(%arg0, %arg1) {dimension = 0 : i64} : (tensor<3x2xf32>, tensor<1x2xf32>) -> tensor<4x2xf32>
  return %0 : tensor<4x2xf32>
}
// CHECK-LABEL: concatenate_float
// CHECK-NOT: tfl.concatenation
// CHECK: stablehlo.concatenate

    }
  }
  return new_arg_type;
}

// Combination of value producer and port of value produced (e.g.,
//   <value result output>:<value in output tensor>,
// so for tf.Const -> tensor<10x20xf32>, [0,2,18] would point to a unique output
// scalar value).
struct ValuePort {
  PointerUnion<Operation*, BlockArgument> producer;
  SmallVector<unsigned int, 2> port;

      tf_device.return %5 : tensor<512xi32>
    }) {device = "/job:localhost/replica:0/task:0/device:CPU:0"} : () -> (tensor<512xi32>, tensor<1024xi32>, tensor<1024xi32>, tensor<1024xf32>)
    // CHECK: %[[COMPILE_OUTPUT:[0-9]*]]:4 = "tf_device.launch"
    // CHECK: "tf._TPUCompileMlir"()
    // CHECK: is_bounded_dynamic_dim: true