^bb0(%arg0: tensor<i32>, %arg1: tensor<i32>, %arg2: tensor<i32>):
    %1 = "mhlo.compare"(%arg0, %arg2) {comparison_direction = #mhlo<comparison_direction LT>} : (tensor<i32>, tensor<i32>) -> tensor<i1>
    "mhlo.return"(%1) : (tensor<i1>) -> ()
  },  {
  ^bb0(%arg0: tensor<i32>, %arg1: tensor<i32>, %arg2: tensor<i32>):
    %1 = mhlo.add %arg0, %arg1 : tensor<i32>
    "mhlo.return"(%1, %arg1, %arg2) : (tensor<i32>, tensor<i32>, tensor<i32>) -> ()

  %21 = "tf.Identity"(%20) {device = ""} : (tensor<i1>) -> tensor<i1>
  %22 = "tf.StridedSlice"(%arg2, %8, %7, %8) {begin_mask = 0 : i64, device = "", ellipsis_mask = 0 : i64, end_mask = 1 : i64, new_axis_mask = 0 : i64, shrink_axis_mask = 0 : i64} : (tensor<?xi64>, tensor<1xi32>, tensor<1xi32>, tensor<1xi32>) -> tensor<?xi64>

  // CHECK: "mhlo.batch_norm_inference"({{.*}}, %arg1, %arg2, %arg3, %arg4) <{epsilon = 1.000000e-03 : f32, feature_index = 3 : i64}> : (tensor<8x8x8x8xf32>, tensor<8xf32>, tensor<8xf32>, tensor<8xf32>, tensor<8xf32>) -> tensor<8x8x8x8xf32>

  %not = "tfl.logical_not"(%arg0) : (tensor<1x2x3x4xi1>) -> tensor<1x2x3x4xi1>
  %sel_v1 = "tfl.select"(%not, %arg1, %arg2) : (tensor<1x2x3x4xi1>, tensor<1x2x3x4xf32>, tensor<1x2x3x4xf32>) -> tensor<1x2x3x4xf32>

	return s.curBlock.NewValue4(s.peekPos(), op, t, arg0, arg1, arg2, arg3)
}

// newValue4I adds a new value with four arguments and an auxint value to the current block.
func (s *state) newValue4I(op ssa.Op, t *types.Type, aux int64, arg0, arg1, arg2, arg3 *ssa.Value) *ssa.Value {
	return s.curBlock.NewValue4I(s.peekPos(), op, t, aux, arg0, arg1, arg2, arg3)
}

//   %4 = mhlo.constant dense<0.000000e+00> : tensor<f32>
//   %5 = "mhlo.broadcast"(%4)
//   %6 = "mhlo.select"(%3, %input, %5)
//   %7 = "mhlo.reduce"(%6, %4) ({
//   ^bb0(%arg1: tensor<f32>, %arg2: tensor<f32>):
//     %9 = mhlo.add %arg1, %arg2 : tensor<f32>
//     "mhlo.return"(%9) : (tensor<f32>) -> ()
//   }) {dimensions = dense<0> : tensor<1xi64>}
//
// If the input's rank N is greater than 2, we will reshape it to R2 first and