double range = std::fabs(max - min);
    if (num_bits <= 8 && range >= 10.0) {
      op.emitWarning()
          << "Tensor range is too wide to be quantized. Use tf.clip_by_value "
             "or tf.relu6 to narrow the tensor range. Range: "
          << range << ", bit width: " << num_bits;
    }
    if (std::abs(max - min) < kNearZeroTolerance) {
      op.emitWarning() << "Tensor range (" << min << ", " << max

    Node* add1 =
        ops::BinaryOp("Add", unstage, b, builder.opts().WithName("add1"));
    Node* relu0 = ops::UnaryOp("Relu", add0, builder.opts().WithName("relu0"));
    ops::UnaryOp("Relu", add1, builder.opts().WithName("relu1"));
    BuildStageNode(builder, "stage", {DT_FLOAT}, {relu0});

    TF_EXPECT_OK(GraphDefBuilderToGraph(builder, graph.get()));
  }

  TF_ASSERT_OK(ClusterScoping(&graph));

     (HasRankAtMost<4> $a),
     (HasRankAtMost<4> $b)]>;
}

// We can eliminate Relu from Relu(SquaredDifference(x, y)),
// since the result of SquaredDifference is always non-negative.
// TFLite interpreter doesn't support Relu+int32 for now. So the test cases
// are failing without the following pattern to optimize Relu away fixes
// the problem.
def OptimizeReluSquaredDifference : Pat<

// CHECK: return %[[CLAMP]] : tensor<?x28x28x16xf32>
// CHECK: }

// -----

// Because the operand of shape_of is other than the target conv,
// should not match conv bias relu6 dynamic pattern.

// CHECK-LABEL: @conv_with_bias_and_relu6_dynamic_shape_not_same_op_fn(
// CHECK-SAME:                    %[[ARG_0:.*]]: tensor<?x28x28x1xf32>

  // CHECK: %0 = "tf.Relu6"(%arg0) {device = "/job:localhost/replica:0/task:0/device:GPU:0"} : (tensor<4xf32>) -> tensor<4xf32>
  // CHECK: return %0
  %cst_6 = arith.constant dense<6.000000e+00> : tensor<f32>
  %0 = "tf.Minimum"(%arg0, %cst_6) : (tensor<4xf32>, tensor<f32>) -> tensor<4xf32>

    Node* add1 =
        ops::BinaryOp("Add", unstage, b, builder.opts().WithName("add1"));
    Node* relu0 = ops::UnaryOp("Relu", add0, builder.opts().WithName("relu0"));
    ops::UnaryOp("Relu", add1, builder.opts().WithName("relu1"));
    MakeStageNode(builder, "stage", {DT_FLOAT}, {relu0});

    return GraphDefBuilderToGraph(builder, graph->get());
  };

// The actual Optimize Pass.
namespace {
#define GEN_PASS_DEF_OPTIMIZEPASS
#include "tensorflow/compiler/mlir/lite/transforms/passes.h.inc"

constexpr char kRelu[] = "RELU";
constexpr char kRelu6[] = "RELU6";
constexpr char kRelu1[] = "RELU_N1_TO_1";

ElementsAttr FlattenTo1D(Attribute a) {
  auto elements = mlir::cast<DenseElementsAttr>(a);

            "TanhGrad", "Pow", "SquaredDifference", "ApproximateEqual",
            // Others
            "AddN", "Bitcast", "Cast", "ClipByValue", "Const", "Empty",
            "Identity", "IdentityN", "Relu", "Relu6", "ReluGrad", "Relu6Grad",
            "LeakyReluGrad", "Elu", "EluGrad", "Selu", "SeluGrad", "Select",
            "SelectV2", "Transpose", "ConjugateTranspose",
            "_UnaryOpsComposition", "CollectiveReduceV2",

  %0 = "tf.SquaredDifference"(%arg0, %arg1) : (tensor<1xf32>, tensor<1xf32>) -> tensor<1xf32>
  %1 = "tf.Relu6"(%0) : (tensor<1xf32>) -> tensor<1xf32>
  func.return %1: tensor<1xf32>

// CHECK-LABEL: squaredDifferenceRelu
// CHECK:  tfl.squared_difference %arg0, %arg1 : tensor<1xf32>
// CHECK:  %1 = "tfl.relu6"(%0) : (tensor<1xf32>) -> tensor<1xf32>
// CHECK:  return
}

  %0 = "tf.Relu"(%arg0) : (tensor<?xui32>) -> tensor<?xui32>
  func.return %0: tensor<?xui32>
}

// -----

// CHECK-LABEL: func @relu6
func.func @relu6(%arg0: tensor<1xi32>) -> tensor<1xi32> {
  // CHECK-DAG: %[[ZERO:.*]] = mhlo.constant dense<0> : tensor<i32>
  // CHECK-DAG: %[[SIX:.*]] = mhlo.constant dense<6> : tensor<i32>