return %2 : tensor<1x3xf32>
  }

  func.func private @composite_dot_general_fn(%arg0: tensor<1x2xf32>, %arg1: tensor<2x3xf32>) -> tensor<1x3xf32> attributes {_from_xla_call_module} {
    %0 = stablehlo.dot_general %arg0, %arg1, contracting_dims = [1] x [0] : (tensor<1x2xf32>, tensor<2x3xf32>) -> tensor<1x3xf32>
    return %0 : tensor<1x3xf32>
  }
}

// CHECK-LABEL: @add_f32(%arg0: tensor<3x3xbf16>, %arg1: tensor<3x3xbf16>) -> tensor<3x3xbf16>
func.func @add_f32(%arg0: tensor<3x3xf32>, %arg1: tensor<3x3xf32>) -> tensor<3x3xf32> {
  // CHECK-NOT: f32
  // CHECK: stablehlo.add
  %0 = stablehlo.add %arg0, %arg1: (tensor<3x3xf32>, tensor<3x3xf32>) -> tensor<3x3xf32>
  return %0 : tensor<3x3xf32>
}

// -----

// CHECK-LABEL: @add_f64(%arg0: tensor<3x3xbf16>, %arg1: tensor<3x3xbf16>) -> tensor<3x3xbf16>

  // CHECK: %[[RESHAPE:.*]] = stablehlo.reshape
  // CHECK-NEXT: return %[[RESHAPE]]

  // CHECK: @main_1
  func.func private @main_1(%arg0: tensor<3x10xf32>) -> tensor<6x5xf32> {
    %0 = stablehlo.reshape %arg0 : (tensor<3x10xf32>) -> tensor<6x5xf32>
    return %0 : tensor<6x5xf32>
  }
  // CHECK: %[[RESHAPE:.*]] = stablehlo.reshape
  // CHECK-NEXT: return %[[RESHAPE]]

  }

  // CHECK-LABEL: pack
  func.func @pack(%arg0: tensor<1xf32>, %arg1: tensor<1xf32>) -> tensor<2x1xf32> {
    %0 = "tfl.pack"(%arg0, %arg1) {axis = 0 : i32, values_count = 2 : i32} : (tensor<1xf32>, tensor<1xf32>) -> tensor<2x1xf32>
    func.return %0 : tensor<2x1xf32>
    // CHECK: %[[VAL_0:.*]] = arith.constant dense<[2, 1]> : tensor<2xi32>

    func.return %1, %0 : tensor<?x5xf32>, tensor<?x5xf32>
  }

  %2 = "tf.Const"() {value = dense<0.0> : tensor<2x2xf32>} : () -> tensor<2x2xf32>
  %3 = "tf.Const"() {value = dense<0.0> : tensor<4x4xf32>} : () -> tensor<4x4xf32>
  "tf.SendTPUEmbeddingGradients"(%2, %3) {_tpu_embedding_layer = "call1", config = "\0A\0B\0C\0D", operandSegmentSizes = array<i32: 2, 0>} : (tensor<2x2xf32>, tensor<4x4xf32>) -> ()

        return %2 : tensor<?x2xf32>
      }
      func.func private @composite_fn_1(%arg0: tensor<?x2xf32>, %arg1: tensor<2x2xf32>, %arg2: tensor<2xf32>) -> tensor<?x2xf32> attributes {_from_xla_call_module, tf_quant.composite_function} {
        return %arg0 : tensor<?x2xf32>
      }
    }
  )mlir";

// CHECK:           %[[VAL_9:.*]] = "tfl.reshape"(%[[VAL_8]], %[[VAL_4]]) {tac.device = "GPU", tac.inference_type = "FLOAT"} : (tensor<2xf32>, tensor<2xi32>) -> tensor<2x1xf32>
// CHECK:           return %[[VAL_9]] : tensor<2x1xf32>
// CHECK:         }

  %0 = "quantfork.stats"(%arg0) {layerStats = dense<[-1.0, 1.0]> : tensor<2xf32>} : (tensor<1x1xf32>) -> tensor<1x1xf32>
  %1 = "tfl.tanh"(%0) : (tensor<1x1xf32>) -> tensor<1x1xf32>
  %2 = "quantfork.stats"(%1) {layerStats = dense<[-1.0, 1.0]> : tensor<2xf32>} : (tensor<1x1xf32>) -> tensor<1x1xf32>
  func.return %2 : tensor<1x1xf32>

  %3 = "tfl.dequantize"(%2) : (tensor<2x2x!quant.uniform<u8:f32, 1.0>>) -> tensor<2x2xf32>
  func.return %3 : tensor<2x2xf32>

// CHECK-NEXT: %[[Q:.*]] = "tfl.quantize"(%arg0) <{qtype = tensor<1x2x!quant.uniform<u8:f32, 1.000000e+00>>}> : (tensor<1x2xf32>) -> tensor<1x2x!quant.uniform<u8:f32, 1.000000e+00>>