%cst = "tf.Const"() {device = "", value = dense<[1]> : tensor<1xi32>} : () -> tensor<1xi32>
  // idx 2 should be a compile time constant
  %0 = "tf.GatherV2"(%arg0, %arg1, %cst) {batch_dims = 1: i64} : (tensor<16x4xf32>, tensor<16xi32>, tensor<1xi32>) -> tensor<16xf32>

  // Just to introduce an extra use for %cst.
  %1 = "tf.AddV2"(%cst, %cst) {device = ""} : (tensor<1xi32>, tensor<1xi32>) -> tensor<1xi32>

  return %0 : tensor<16xf32>
}

  %2 = "tfl.shape"(%1) : (tensor<?xf32>) -> tensor<1xi32>
  func.return %2, %2 : tensor<1xi32>, tensor<1xi32>

// CHECK: %[[s1:.*]] = "tfl.shape"(%arg0) : (tensor<?x!quant.uniform<u8:f32, 1.000000e+00>>) -> tensor<1xi32>
// CHECK-NEXT: %[[s1]], %[[s1]] : tensor<1xi32>, tensor<1xi32>
}

// CHECK-LABEL: NotQuantizePow
func.func @NotQuantizePow(%arg0: tensor<4x!quant.uniform<u8:f32, 1.0>>,

  func.return %0 : tensor<!tf_type.variant<tensor<*xi32>>>
}

// -----

// CHECK-LABEL: listGetItem
func.func @listGetItem(%arg0: tensor<!tf_type.variant<tensor<*xi32>>>, %arg1: tensor<i32>, %arg2: tensor<2xi32>) -> tensor<2xi32> {

  // CHECK-NEXT: %[[CONST1:.*]] = "tf.Const"() <{value = dense<1> : tensor<1xi32>}> : () -> tensor<1xi32>
  // CHECK-NEXT: %[[NEW_SIZE:.*]] = "tf.AddV2"(%[[READ_SIZE]], %[[CONST1]]) : (tensor<1xi32>, tensor<1xi32>) -> tensor<1xi32>
  // CHECK-NEXT: "tf.AssignVariableOp"(%[[SIZE]], %[[NEW_SIZE]]) : (tensor<!tf_type.resource<tensor<1xi32>>>, tensor<1xi32>) -> ()

    %1 = "tfl.abs"(%arg2) {fused_activation_function = "NONE"} : (tensor<5xf32>) -> tensor<5xf32>
    %2 = stablehlo.add %1, %arg2 : tensor<5xf32>
    %3 = "tfl.abs"(%2) {fused_activation_function = "NONE"} : (tensor<5xf32>) -> tensor<5xf32>
    "stablehlo.return"(%3) : (tensor<5xf32>) -> ()
  }) {dimensions = array<i64: 0>} : (tensor<7x5xf32>, tensor<5xf32>) -> tensor<5xf32>
  func.return %0: tensor<5xf32>
}

// -----

    %s0 = "tf.Const"() {value = dense<[501, 1, 32, 1280]> : tensor<4xi32>} : () -> tensor<4xi32>
    %s1 = "tf.Const"() {value = dense<[  1, 1,  1, 1280]> : tensor<4xi32>} : () -> tensor<4xi32>

    // CHECK: "tf.BroadcastGradientArgs"
    // CHECK-NOT: _xla_outside_compilation
    %r0, %r1 = "tf.BroadcastGradientArgs"(%s0, %s1) {} : (tensor<4xi32>, tensor<4xi32>) -> (tensor<1xi32>, tensor<3xi32>)
    tf_device.return

  %m0 = mhlo.constant dense<0> : tensor<i32>
  %t0 = "tf.Const"() { value = dense<0> : tensor<1xi32> } : () -> tensor<1xi32>
  %m1 = mhlo.constant dense<1> : tensor<i32>
  %t1 = "tf.Const"() { value = dense<1> : tensor<1xi32> } : () -> tensor<1xi32>
  %m2 = mhlo.constant dense<1> : tensor<i32>
  %t2 = "tf.Const"() { value = dense<1> : tensor<1xi32> } : () -> tensor<1xi32>
  %m3 = mhlo.constant dense<1> : tensor<i32>

// PerChannel-DAG: %[[zp1:.*]] = "tf.Const"() <{value = dense<{{[0-9]+}}> : tensor<3xi32>}> : () -> tensor<3xi32>
// PerChannel-DAG: %[[zp2:.*]] = "tf.Const"() <{value = dense<{{[0-9]+}}> : tensor<6xi32>}> : () -> tensor<6xi32>
// PerChannel: %[[out_1:.*]] = "tf.PartitionedCall"(%arg0, %[[q_w1]], %[[scale1]], %[[zp1]]) <{config = "", config_proto = "", executor_type = "",

    %2 = stablehlo.reshape %1 : (tensor<i32>) -> tensor<1xi32>
    %3 = stablehlo.concatenate %2, %cst_0, %cst_1, %cst_2, dim = 0 : (tensor<1xi32>, tensor<1xi32>, tensor<1xi32>, tensor<1xi32>) -> tensor<4xi32>
    %4 = stablehlo.dynamic_broadcast_in_dim %arg2, %3, dims = [0, 1, 2, 3] : (tensor<1x1x1x2xf32>, tensor<4xi32>) -> tensor<?x3x4x2xf32>
    %5 = stablehlo.add %0, %4 : tensor<?x3x4x2xf32>

  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"