// CHECK-DAG: %[[MATMUL_LHS_SHAPE:.*]] = "tf.Const"() <{value = dense<[4, 5]> : tensor<2xi64>}> : () -> tensor<2xi64>
  // CHECK-DAG: %[[MATMUL_RHS_SHAPE:.*]] = "tf.Const"() <{value = dense<[5, 6]> : tensor<2xi64>}> : () -> tensor<2xi64>

  // CHECK: %[[LHS_1:.*]] = "tf.Reshape"(%arg0, %[[MATMUL_LHS_SHAPE]]) : (tensor<1x4x5xf32>, tensor<2xi64>) -> tensor<4x5xf32>

  %2 = "mhlo.dynamic_broadcast_in_dim"(%cst_0, %1) <{broadcast_dimensions = dense<3> : tensor<1xi64>}> : (tensor<2xf32>, tensor<4xindex>) -> tensor<?x256x256x2xf32>
  %3 = mhlo.multiply %0, %2 : tensor<?x256x256x2xf32>
  %4 = "mhlo.dynamic_broadcast_in_dim"(%cst_1, %1) <{broadcast_dimensions = dense<3> : tensor<1xi64>}> : (tensor<2xf32>, tensor<4xindex>) -> tensor<?x256x256x2xf32>
  %5 = mhlo.add %3, %4 : tensor<?x256x256x2xf32>

    %shape_64 = "tf.TensorListElementShape"(%tl_1) : (tensor<!tf_type.variant<tensor<?x1xf32>>>) -> tensor<?xi64>
    // CHECK: %[[CAST:.*]] = "tf.Cast"(%[[ELEMENT_SHAPE]]){{.*}}: (tensor<2xi32>) -> tensor<2xi64>
    // CHECK: "tf._SomeOtherOp"(%[[ELEMENT_SHAPE]], %[[CAST]])
    "tf._SomeOtherOp"(%shape_32, %shape_64) : (tensor<?xi32>, tensor<?xi64>) -> ()
    func.return
  }

  // CHECK-LABEL: refine_pop_back_results_from_operands

  %0 = "tf.Const"() {value = dense<42> : tensor<1xi64>} : () -> tensor<1xi64>
  return
}
func.func @init2() attributes {tf_saved_model.exported_names = ["init2"]} {
  %0 = "tf.Const"() {value = dense<43> : tensor<1xi64>} : () -> tensor<1xi64>
  return
}
func.func @init3() attributes {tf_saved_model.exported_names = ["init3"]} {
  %0 = "tf.Const"() {value = dense<44> : tensor<1xi64>} : () -> tensor<1xi64>
  return
}

module {
  func.func @main() -> tensor<1xi64> {
  %0 = "tfl.custom"() {custom_code = "stablehlo.constant", custom_option = #tfl<const_bytes : "0x76616C75650001020109010101062C022401">} : () -> tensor<1xi64>
  func.return %0 : tensor<1xi64>
  }
}

// CHECK:  module {
// CHECK-NEXT:    func @main() -> tensor<1xi64> {
// CHECK-NEXT:    %[[c0:.+]] = stablehlo.constant dense<2> : tensor<1xi64>
// CHECK-NEXT:    return %[[c0]] : tensor<1xi64>

  %cst = arith.constant dense<[2, 100]> : tensor<2xi64>
  // CHECK: tac.cost = 4.000000e+01
  %0 = "tfl.concatenation"(%arg0, %arg1) {axis = 0 : i32, fused_activation_function = "NONE", tac.device = "GPU"} : (tensor<100xf32>, tensor<100xf32>) -> tensor<200xf32>
  // CHECK: tac.cost = 4.040000e+01
  %1 = "tfl.reshape"(%0, %cst) {tac.device = "GPU"} : (tensor<200xf32>, tensor<2xi64>) -> tensor<2x100xf32>
  func.return %1 : tensor<2x100xf32>
}

  // CHECK-DAG: %[[SCALE_BCAST:.+]] = "mhlo.broadcast_in_dim"(%[[SCALE]]) <{broadcast_dimensions = dense<1> : tensor<1xi64>}> : (tensor<256xf32>) -> tensor<4x256xf32>
  // CHECK-DAG: %[[OFFSET_BCAST:.+]] = "mhlo.broadcast_in_dim"(%[[OFFSET]]) <{broadcast_dimensions = dense<1> : tensor<1xi64>}> : (tensor<256xf32>) -> tensor<4x256xf32>

// HISTOGRAM-PERCENTILE-CHECK-NEXT:  [[lhs:%.*]], {{.*}}, {{.*}}, {{.*}} = "tf.CustomAggregator"(%arg0) <{calibration_method = 3 : i32, id = "composite_conv2d_with_relu6_fn_arg_0_calibration_method_3", max_percentile = 9.999900e+01 : f32, min_percentile = 1.000000e-03 : f32, num_bins = 512 : i32}> : (tensor<*xf32>) -> (tensor<*xf32>, tensor<f32>, tensor<f32>, tensor<512xi64>)

// RUN: tf-opt %s -tfl-prepare-tf -tfl-legalize-tf='run-tfl-runtime-verification=false'  -tfl-optimize | FileCheck %s

func.func @broadcast_to_bf16(%arg0: tensor<3xbf16>, %arg1: tensor<2xi64>) -> tensor<3x3xbf16> {
  %0 = "tf.BroadcastTo"(%arg0, %arg1) : (tensor<3xbf16>, tensor<2xi64>) -> tensor<3x3xbf16>
  func.return %0: tensor<3x3xbf16>

// CHECK-LABEL: broadcast_to_bf16
// CHECK:  [[CST:%.*]] = arith.constant dense<1.000000e+00> : tensor<3x3xbf16>

    %0 = stablehlo.constant dense<500> : tensor<100xi64>
    %1 = stablehlo.constant dense<0> : tensor<100xi64>
    %2 = stablehlo.compare  LT, %arg2, %1 : (tensor<100xi64>, tensor<100xi64>) -> tensor<100xi1>
    %3 = stablehlo.add %arg2, %0 : tensor<100xi64>
    %4 = stablehlo.select %2, %3, %arg2 : tensor<100xi1>, tensor<100xi64>
    %5 = stablehlo.reshape %4 : (tensor<100xi64>) -> tensor<100x1xi64>