%2 = "tf.Cast"(%1) {Truncate = false} : (tensor<1x2xbf16>) -> tensor<1x2xf32>
  %3 = "tf.IdentityN"(%2) {device = ""} : (tensor<1x2xf32>) -> tensor<1x2xf32>
  return %3 : tensor<1x2xf32>
}

// CHECK: func @cast_bf16_matmul_to_fp32
// CHECK-DAG: %[[cst:.*]] = "tf.Const"() <{value = dense<{{.*}}> : tensor<10x2xf32>}> : () -> tensor<10x2xf32>
// CHECK: %[[matmul:.*]] = "tf.MatMul"(%arg0, %[[cst]])

}

// -----

// CHECK-LABEL: @acos
func.func @acos(%arg0: tensor<2xf32>) -> tensor<2xf32> {
  // CHECK-NOT:  tf.Acos
  %0 = "tf.Acos"(%arg0) : (tensor<2xf32>) -> tensor<2xf32>
  func.return %0 : tensor<2xf32>
}

// -----

// NOFALLBACK-LABEL: @xla_svd
func.func @xla_svd(%arg0: tensor<1x1xf32>) -> (tensor<1xf32>, tensor<1x1xf32>, tensor<1x1xf32>) {
  // NOFALLBACK: XlaSvd

// `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

// -----

  // If operand has the same shape as a result, we can fold it.
  %3 = "tf.AddV2"(%arg1, %0) : (tensor<4x2xf32>, tensor<1x2xf32>) -> tensor<4x2xf32>
  %4 = "tf.AddV2"(%0, %arg1) : (tensor<1x2xf32>, tensor<4x2xf32>) -> tensor<4x2xf32>

  // CHECK: %[[CONST:.*]] = "tf.Const"()
  // CHECK-DAG: %[[ADD1:.*]] = "tf.AddV2"(%arg0, %[[CONST]])

    }) {is_stateless = false} : (tensor<i32>, tensor<!tf_type.variant<tensor<?x1xf32>>>) -> (tensor<i32>, tensor<!tf_type.variant<tensor<?x1xf32>>>)
    %elem_1 = "tf._SomeOtherOp"() : () -> tensor<8x1xf32>
    %tl_set_item = "tf.TensorListSetItem"(%while#1, %one, %elem_1) : (tensor<!tf_type.variant<tensor<?x1xf32>>>, tensor<i32>, tensor<8x1xf32>) -> tensor<!tf_type.variant<tensor<?x1xf32>>>
    func.return
  }

  %outputs_1, %outputs_2 = "tf.IdentityN"(%outputs, %outputs_0) {device = "TPU"} : (tensor<2x2xf32>, tensor<2x2xf32>) -> (tensor<2x2xf32>, tensor<2x2xf32>)
  func.return %outputs_0, %outputs_1 : tensor<2x2xf32>, tensor<2x2xf32>
}

// CHECK-LABEL: func @keep_identity_n_test
func.func @keep_identity_n_test(%arg0: tensor<2x2xf32>, %arg1: tensor<2x2xf32>) -> (tensor<2x2xf32>, tensor<2x2xf32>) {
    // CHECK: tf.MatMul

    %9 = "tfl.fully_connected"(%8, %cst_10, %cst_0) {fused_activation_function = "NONE", keep_num_dims = false, weights_format = "DEFAULT"} : (tensor<4x5xf32>, tensor<8x5xf32>, tensor<8xf32>) -> tensor<4x8xf32>
    %10:4 = "tfl.split"(%cst_5, %9) {num_splits = 4 : i32} : (tensor<i32>, tensor<4x8xf32>) -> (tensor<4x2xf32>, tensor<4x2xf32>, tensor<4x2xf32>, tensor<4x2xf32>)

  func.return %24 : tensor<1x4xf32>
}

    %0 = "tf.MatMul"(%arg0, %arg1) {attr_map = "0:transpose_a,1:transpose_b", device = "", transpose_a = false, transpose_b = false} : (tensor<2x2xf32>, tensor<2x2xf32>) -> tensor<2x2xf32>
    return %0 : tensor<2x2xf32>
  }

      %2 = "quantfork.stats"(%1) {layerStats = dense<[5.00000000e-6, 7.00000000e-1]> : tensor<2xf32>} : (tensor<1x3xf32>) -> tensor<1x3xf32>
      return %2 : tensor<1x3xf32>
    }

    func.func private @composite_dot_general_fn(%arg0: tensor<1x2xf32>, %arg1: tensor<2x3xf32>) -> tensor<1x3xf32> attributes {_from_xla_call_module} {