module {
    func.func @main(%arg0: tensor<2x2x2xf32>, %arg1: tensor<2x2x2xf32>) -> tensor<2x2x2xf32> attributes {_from_xla_call_module} {
      %0 = stablehlo.dot_general %arg0, %arg1,
        batching_dims = [0] x [0],
        contracting_dims = [2] x [1],
        precision = [DEFAULT, DEFAULT]
      : (tensor<2x2x2xf32>, tensor<2x2x2xf32>) -> tensor<2x2x2xf32>
        return %0 : tensor<2x2x2xf32>
    }
  }
)mlir";

      return %0 : tensor<1x1x4xf32>
    }
  )mlir";

  OwningOpRef<ModuleOp> module_op =
      ParseModuleOpString(kModuleXlaCallModuleOpWithDefaultQuantizationMethod);
  ASSERT_TRUE(module_op);

      return %0 : tensor<1x1x4xf32>
    }
  )mlir";

  const OwningOpRef<ModuleOp> module_op =
      ParseModuleOpString(kXlaCallModuleOpWithQuantizationMethodAttr);
  ASSERT_TRUE(module_op);

  func::FuncOp main_fn = FindMainFuncOp(*module_op);

  %1 = "tfl.add"(%cst_2, %cst_1) {fused_activation_function = "NONE"} : (tensor<2x2x2xf32>, tensor<  2x2xf32>) -> tensor<2x2x2xf32>
  %2 = "tfl.add"(%cst_0, %cst_2) {fused_activation_function = "NONE"} : (tensor<    2xf32>, tensor<2x2x2xf32>) -> tensor<2x2x2xf32>

  func.return %0, %1, %2 : tensor<2x2xf32>, tensor<2x2x2xf32>, tensor<2x2x2xf32>

    %unshaped = "tf.Cast"(%arg1) : (tensor<1x2x3xf32>) -> tensor<*xf32>
    // CHECK: <{is_stateless = true}>
    %0 = "tf.IfRegion"(%arg0) <{is_stateless = true}> ({
      // CHECK: "tf.Add"{{.+}}(tensor<1x2x3xf32>, tensor<1x2x3xf32>) -> tensor<1x2x3xf32>
      // CHECK: "tf.Yield"{{.+}}(tensor<1x2x3xf32>) -> ()

}
func.func @_func(%arg0: tensor<2x4xf32>, %arg1: tensor<4x2xf32>) -> tensor<2x2xf32> {
  %0 = "tf.MatMul"(%arg0, %arg1) {_XlaSharding = "\08\03\1A\02\02\01\22\02\00\01"} : (tensor<2x4xf32>, tensor<4x2xf32>) -> tensor<2x2xf32>
  %1 = "tf.Identity"(%0) : (tensor<2x2xf32>) -> tensor<2x2xf32>
  return %1 : tensor<2x2xf32>
}

// -----
// The following op sharding is used in the following test case:

  %0 = "tfl.var_handle"() {container = "", shared_name = "states"} : () -> tensor<!tf_type.resource<tensor<1x2x3xf32>>>
  %1 = "tfl.read_variable"(%0) : (tensor<!tf_type.resource<tensor<1x2x3xf32>>>) -> tensor<1x2x3xf32>
  %2 = "tfl.quantize"(%1) {qtype = tensor<1x2x3x!quant.uniform<u8<1:255>:f32, 0.047244094488188976:128>>} : (tensor<1x2x3xf32>) -> tensor<1x2x3x!quant.uniform<u8<1:255>:f32, 0.047244094488188976:128>>

// CHECK: stablehlo.maximum
// CHECK: (tensor<1x5x5x4xf32>, tensor<f32>) -> tensor<1x2x2x4xf32>
// CHECK: %[[TRANSPOSE_1:.+]] = stablehlo.transpose %[[REDUCE_WINDOW_MAX]], dims = [0, 3, 1, 2] : (tensor<1x2x2x4xf32>) -> tensor<1x4x2x2xf32>
// CHECK: return %[[TRANSPOSE_1]]

// -----

// Tests that a `maximum(add(convolution(%activation, %weight), %bias), %zero)`

  // CHECK: [[Result:%.*]] = "tf.Case"(%arg0, %arg1)
  // CHECK-SAME: branches = [@tf.CaseRegion_branch0{{.*}}, @tf.CaseRegion_branch1{{.*}}]
  // CHECK-SAME: is_stateless = false
  %0 = "tf.CaseRegion"(%arg0) ({
    %1 = "tf.Cast"(%arg1) {Truncate = false} : (tensor<!tf_type.resource<tensor<1x2x3xf32>>>) -> tensor<!tf_type.resource>

func.func @ReorderAddWithConstant(%arg0: tensor<2x2xf32>) -> tensor<2x2xf32> {
  %cst = arith.constant dense<1.0> : tensor<2x2xf32>
  %cst_1 = arith.constant dense<2.0> : tensor<2x2xf32>
  %0 = "tfl.add"(%arg0, %cst) {fused_activation_function = "NONE"} : (tensor<2x2xf32>, tensor<2x2xf32>) -> tensor<2x2xf32>
  %1 = "tfl.add"(%0, %cst_1) {fused_activation_function = "NONE"} : (tensor<2x2xf32>, tensor<2x2xf32>) -> tensor<2x2xf32>