rewriter.replaceOp(slice_op, output_reshape_op.getResult());

  return success();
}

// ================== fully_connected ========================

// TFL fully_connected basically does:
// Weight * Input + bias.
// Input layout is : [..., depth]
// Weight layout is : [output, depth]
// Bias is [output].
//
// While conv2d is:
// Filter: [NHWC]

// CHECK-NOT: "tfl.reshape"
// CHECK: "tfl.slice"

// -----

func.func @fullyConnectedToConv(%arg0: tensor<384x384xf32>, %arg1: tensor<512x384xf32>, %arg2: tensor<512xf32>) -> tensor<384x512xf32> {
  %0 = "tfl.fully_connected"(%arg0, %arg1, %arg2) {fused_activation_function = "NONE", keep_num_dims = false, weights_format = "DEFAULT"} : (tensor<384x384xf32>, tensor<512x384xf32>, tensor<512xf32>) -> tensor<384x512xf32>
  func.return %0: tensor<384x512xf32>

// An implementation of TensorFlow fully_connected (a.k.a Dense) layer.
table FullyConnectedOptions {
  // Parameters for FullyConnected version 1 or above.
  fused_activation_function:ActivationFunctionType;

  // Parameters for FullyConnected version 2 or above.
  weights_format:FullyConnectedOptionsWeightsFormat = DEFAULT;

  // Parameters for FullyConnected version 5 or above.

    %8 = "tfl.concatenation"(%7, %arg5) {axis = 1 : i32, fused_activation_function = "NONE"} : (tensor<4x3xf32>, tensor<4x2xf32>) -> tensor<4x5xf32>
    %9 = "tfl.fully_connected"(%8, %cst_10, %cst_0) {fused_activation_function = "NONE", keep_num_dims = false, weights_format = "DEFAULT"} : (tensor<4x5xf32>, tensor<8x5xf32>, tensor<8xf32>) -> tensor<4x8xf32>

        // input activations need to be asymmetrically quantized. Refer to the
        // implementation for fully-connected as an example in
        // tensorflow/lite/kernels/fully_connected.cc. The kernels will handle
        // the asymmetric_quantize_inputs attribute in the builtin option.
        dynamic_range_quantized_user->setAttr(
            kAsymmetricQuantizeInputsAttr,

  %4 = "tfl.dequantize"(%3) : (tensor<32x12x!quant.uniform<u8<1:255>:f32, 0.021826678373682216:151>>) -> tensor<32x12xf32>
  %5 = "tfl.fully_connected"(%2, %4, %cst) {fused_activation_function = "NONE", keep_num_dims = false, weights_format = "DEFAULT"} : (tensor<1x224x224x3xf32>, tensor<32x12xf32>, tensor<32xf32>) -> tensor<1x112x112x32xf32>

      return dequantize.getInput();
    } else {
      return value;
    }
  };

  // Returns true if the TF::BatchMatMul operation can be converted to
  // tfl.fully_connected.
  auto can_convert_to_fully_connected =
      [&](BatchMatMulOpType& batch_matmul_op) {
        Value input_rhs = get_real_input_value(batch_matmul_op.getY());

        DenseElementsAttr constant;

  func.return %0, %1, %2 : tensor<2xi32>, tensor<2xf32>, tensor<i32>
}

// -----

func.func @fully_connected(%arg0: tensor<1x37xf32>, %arg1: tensor<40x37xf32>, %arg2: tensor<40xf32>) -> tensor<1x40xf32> {

// CHECK: %[[ARG:.*]] = "tfl.transpose"(%arg1, %[[CST]]) : (tensor<37x40xf32>, tensor<2xi32>) -> tensor<40x37xf32>
// CHECK: %[[CST_0:.*]] = "tfl.no_value"() <{value}> : () -> none
// CHECK: "tfl.fully_connected"(%arg0, %[[ARG]], %[[CST_0]]) <{fused_activation_function = "NONE", keep_num_dims = false, weights_format = "DEFAULT"}> : (tensor<40x37xf32>, tensor<40x37xf32>, none) -> tensor<40x40xf32>
}

    // [b, 0, 1, f]x[0, 1, i, o]->[b, 0, 1, f]
    // `stablehlo.dot_general` can take various formats. We only per-channel
    // quantize non-batch ops.
    // `stablehlo.dot_general` legalizable to `tfl.fully_connected` has a
    // filter rank of 2 with the last dimension as the channel dimension.
    const int64_t quantization_dimension =
        mlir::cast<ShapedType>(filter_type).getShape().size() - 1;