// CHECK:           [[VAL_7:%.*]] = "tfl.slice"([[VAL_1]], [[VAL_3]], [[VAL_5]]) : (tensor<1x8x8x1024xf32>, tensor<4xi32>, tensor<4xi32>) -> tensor<1x8x8x256xf32>
// CHECK:           [[VAL_8:%.*]] = "tfl.slice"([[VAL_1]], [[VAL_4]], [[VAL_5]]) : (tensor<1x8x8x1024xf32>, tensor<4xi32>, tensor<4xi32>) -> tensor<1x8x8x256xf32>

  %0 = "tf.RealDiv"(%arg0, %arg0) : (tensor<2xi32>, tensor<2xi32>) -> tensor<2xi32>
  func.return %0: tensor<2xi32>
}

// CHECK-LABEL: func @sub
func.func @sub(%arg0: tensor<2xi32>) -> tensor<2xi32> {
  // CHECK-NEXT:  %0 = mhlo.subtract %arg0, %arg0 : tensor<2xi32>
  // CHECK-NEXT:  return %0 : tensor<2xi32>
  %0 = "tf.Sub"(%arg0, %arg0) : (tensor<2xi32>, tensor<2xi32>) -> tensor<2xi32>
  func.return %0: tensor<2xi32>
}

// CHECK-DAG:       %[[VAL_3:.*]] = "tfl.pseudo_const"(){{.*}}dense<2> : tensor<1xi32>
// CHECK-DAG:       %[[VAL_4:.*]] = "tfl.pseudo_const"(){{.*}}dense<[2, 1]> : tensor<2xi32>
// CHECK:           %[[VAL_5:.*]] = "tfl.reshape"(%[[VAL_0]], %[[VAL_2]]) {tac.device = "GPU", tac.inference_type = "FLOAT"} : (tensor<1xf32>, tensor<4xi32>) -> tensor<1x1x1x1xf32>

#loc3])\0Amodule {\0A  func.func @host_func(%arg0: tensor<3360x?xi32> loc(fused[#loc1, #loc2, #loc3]), %arg1: tensor<3xi32> loc(fused[#loc1, #loc2, #loc3]), %arg2: tensor<i32> loc(fused[#loc1, #loc2, #loc3]), %arg3: tensor<2xi32> loc(fused[#loc1, #loc2, #loc3])) -> (tensor<1x1120x?xi32>, tensor<1x1120x?xi32>, tensor<1120x?xi32>, tensor<2xi32>) {\0A    %0 = \22tf.Reshape\22(%arg0, %arg1) {_xla_outside_compilation = \220\22} : (tensor<3360x?xi32>, tensor<3xi32>) -> tensor<3x1120x?xi32> loc(#loc9)\0A    %1:3 = ...

    %4 = "tf.Const"() {value = dense<> : tensor<0xi32>} : () -> tensor<0xi32>
    %5 = "tf.Const"() {value = dense<2.500000e-01> : tensor<f32>} : () -> tensor<f32>
    %6 = "tf.Const"() {value = dense<0> : tensor<1xi32>} : () -> tensor<1xi32>
    %7 = "tf.Const"() {value = dense<[-1, 1001]> : tensor<2xi32>} : () -> tensor<2xi32>
    %8 = "tf.Const"() {value = dense<[1, 2]> : tensor<2xi32>} : () -> tensor<2xi32>

    %26 = "tfl.slice"(%arg3, %22, %25) : (tensor<*xf32>, tensor<?xi32>, tensor<?xi32>) -> tensor<*xf32>
    %27 = "tfl.reshape"(%5, %cst_12) : (tensor<i32>, tensor<1xi32>) -> tensor<1xi32>
    %28 = "tfl.concatenation"(%27, %21) {axis = 0 : i32, fused_activation_function = "NONE"} : (tensor<1xi32>, tensor<?xi32>) -> tensor<?xi32>

  %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>>,

  // 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>) -> ()

    %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