%13 = stablehlo.subtract %10, %12 : tensor<1x4x3xf32>  // Precalculated zp_neg.
    %14 = stablehlo.broadcast_in_dim %4, dims = [0, 1, 2] : (tensor<1x1x3xf32>) -> tensor<1x4x3xf32>  // Optional
    %15 = stablehlo.multiply %13, %14 : tensor<1x4x3xf32>  // s1 * s2
    %16 = call @uniform_quantize_1(%15, %5, %6) : (tensor<1x4x3xf32>, tensor<1x1x1xf32>, tensor<1x1x1xi8>) -> tensor<1x4x3xi8>

  }) {is_stateless = true} : (tensor<i32>, tensor<i32>, tensor<4x4x2xf32>, tensor<4x2xf32>, tensor<4x2xf32>, tensor<4x4x3xf32>) -> (tensor<i32>, tensor<i32>, tensor<*xf32>, tensor<4x2xf32>, tensor<4x2xf32>, tensor<*xf32>)
  %2 = "tfl.shape"(%1#2) : (tensor<*xf32>) -> tensor<?xi32>
  %3 = "tfl.reshape"(%1#2, %2) : (tensor<*xf32>, tensor<?xi32>) -> tensor<?x4x2xf32>

    // CHECK: %[[RESHAPE1:.*]] = mhlo.reshape %arg1 : (tensor<4x1x4xf32>) -> tensor<4x4xf32>
    // CHECK: %[[UPDATED_ARG1:.*]] = "mhlo.broadcast_in_dim"(%[[RESHAPE1]]) <{broadcast_dimensions = dense<[0, 2]> : tensor<2xi64>}> : (tensor<4x4xf32>) -> tensor<4x4x4xf32>

    // CHECK: %[[RESULT:.*]] = mhlo.atan2 %[[UPDATED_ARG0]], %[[UPDATED_ARG1]] : tensor<4x4x4xf32>

    %5 = "quantfork.qcast"(%4) {volatile} : (tensor<3x4x2xf32>) -> tensor<3x4x2x!quant.uniform<i8:f32, 0.13170163023705575:-1>>
    %6 = "quantfork.dcast"(%5) : (tensor<3x4x2x!quant.uniform<i8:f32, 0.13170163023705575:-1>>) -> tensor<3x4x2xf32>
    %7 = "stablehlo.gather"(%6, %arg2) {
    dimension_numbers = #stablehlo.gather<

    func.func @main() -> (tensor<1x4x4xf32>) {
      %%arg0 = "tf.Const"() {value = dense<-3.0> : tensor<1x4x2xf32>} : () -> tensor<1x4x2xf32>
      %%arg1 = "tf.Const"() {value = dense<-3.0> : tensor<1x2x4xf32>} : () -> tensor<1x2x4xf32>

      %%1 = "tf.%s"(%%arg0, %%arg1) {T = f32, adj_x = false, adj_y = false, grad_x = false, grad_y = false, device = ""} : (tensor<1x4x2xf32>, tensor<1x2x4xf32>) -> tensor<1x4x4xf32>

func.func @reshape_removeIdentity(tensor<4x4x4xf32>) -> tensor<4x4x4xf32> {
^bb0(%arg0: tensor<4x4x4xf32>) :
  %cst = arith.constant dense<[4, 4, 4]> : tensor<3xi32>
  %0 = "tfl.reshape"(%arg0, %cst) : (tensor<4x4x4xf32>, tensor<3xi32>) -> tensor<4x4x4xf32>
  func.return %0 : tensor<4x4x4xf32>

// CHECK-LABEL: func @reshape_removeIdentity
// CHECK:  return %arg0 : tensor<4x4x4xf32>
}

    %add_out, %add_control = tf_executor.island wraps "tf.AddV2"(%const_out, %const_out) {_xla_compile_device_type = "TPU", _replication_info = "cluster"} : (tensor<4x4xf32>, tensor<4x4xf32>) -> tensor<4x4xf32>
    %replicated_out, %replicated_control = tf_executor.island wraps "tf.TPUReplicatedOutput"(%add_out) : (tensor<4x4xf32>) -> (tensor<4x4xf32>)

func.func @lower_einsum(%arg0: tensor<3x4x5xf32>, %arg1: tensor<3x5x6xf32>) -> tensor<3x4x6xf32> {
  %0 = "tf.Einsum"(%arg0, %arg1) {T = "tfdtype$DT_FLOAT", equation = "ijk,ikm->ijm"}: (tensor<3x4x5xf32>, tensor<3x5x6xf32>) -> tensor<3x4x6xf32>
  func.return %0 : tensor<3x4x6xf32>
}
// CHECK-LABEL: lower_einsum
// CHECK: "tf.BatchMatMulV2"(%arg0, %arg1) <{adj_x = false, adj_y = false}> : (tensor<3x4x5xf32>, tensor<3x5x6xf32>) -> tensor<3x4x6xf32>

func.func @lift_float_batch_matmul(%arg0: tensor<4x4x3xf32>) -> (tensor<4x4x3xf32>) {
  %cst = "tf.Const"() {device = "", value = dense<1.0> : tensor<4x3x3xf32>} : () -> tensor<4x3x3xf32>
  %0 = "tf.BatchMatMulV2"(%arg0, %cst) {adj_x = false, adj_y = false, device = ""} : (tensor<4x4x3xf32>, tensor<4x3x3xf32>) -> tensor<4x4x3xf32>
  return %0 : tensor<4x4x3xf32>

func.func @main(tensor<4x4x4xf32>, tensor<4x4xf32>, tensor<4x4xf32>, tensor<4x4xf32>, tensor<4x4xf32>, tensor<4x4xf32>, tensor<4x4xf32>, tensor<4x4xf32>, tensor<4x4xf32>, tensor<4xf32>, tensor<4xf32>, tensor<4xf32>, tensor<4xf32>, tensor<4xf32>, tensor<4xf32>, tensor<4xf32>, tensor<4x4xf32>, tensor<4xf32>, tensor<4x4xf32>, tensor<4x4xf32>, tensor<4x4xf32>, tensor<4x4xf32>) -> tensor<4x4x4xf32> {
// CHECK: {
// CHECK-NEXT:   version: 3,