%2 = "tfl.add"(%cst_0, %cst_2) {fused_activation_function = "NONE"} : (tensor<    2xi32>, tensor<2x2x2xi32>) -> tensor<2x2x2xi32>

  func.return %0, %1, %2 : tensor<2x2xi32>, tensor<2x2x2xi32>, tensor<2x2x2xi32>

// CHECK-DAG:  %[[CST:.*]] = arith.constant dense<{{\[\[}}11, 22], [13, 24]]> : tensor<2x2xi32>

      %%arg0 = "tf.Const"() {value = dense<-3.0> : tensor<1x4x2xf32>} : () -> tensor<1x4x2xf32>
      %%arg1 = "tf.Const"() {value = dense<-3.0> : tensor<1x2x4xf32>} : () -> tensor<1x2x4xf32>

      %%1 = "tf.%s"(%%arg0, %%arg1) {T = f32, adj_x = false, adj_y = false, grad_x = false, grad_y = false, device = ""} : (tensor<1x4x2xf32>, tensor<1x2x4xf32>) -> tensor<1x4x4xf32>

      func.return %%1 : tensor<1x4x4xf32>
    }

    %8 = stablehlo.convert %7 : (tensor<1x4x2xi8>) -> tensor<1x4x2xf32>
    %9 = stablehlo.convert %2 : (tensor<2x3xi8>) -> tensor<2x3xf32>
    %10 = stablehlo.dot_general %8, %9, contracting_dims = [2] x [0] : (tensor<1x4x2xf32>, tensor<2x3xf32>) -> tensor<1x4x3xf32>
    %11 = stablehlo.convert %3 : (tensor<1x1x3xi32>) -> tensor<1x1x3xf32>
    %12 = stablehlo.broadcast_in_dim %11, dims = [0, 1, 2] : (tensor<1x1x3xf32>) -> tensor<1x4x3xf32>  // Optional

  
  %40 = "tfl.read_variable"(%4) : (tensor<!tf_type.resource>) -> tensor<1x2x3xf32>
  %41 = "quantfork.stats"(%40) {layerStats = dense<[0.0, 1.0]> : tensor<2xf32>} : (tensor<1x2x3xf32>) -> tensor<1x2x3xf32>
  %42 = "tfl.concatenation"(%41, %0) {axis = 1 : i32, fused_activation_function = "NONE"} : (tensor<1x2x3xf32>, tensor<1x2x3xf32>) -> tensor<1x4x3xf32>

// CHECK-LABEL: QuantizeUnidirectionalLstmFullPerTensor
func.func @QuantizeUnidirectionalLstmFullPerTensor(%arg0: tensor<1x2x3xf32>) -> (tensor<1x2x3xf32>) {
  %input = "quantfork.stats"(%arg0) {layerStats = dense<[0.0, 1.0]> : tensor<2xf32>} : (tensor<1x2x3xf32>) -> tensor<1x2x3xf32>
  %1 = "tfl.pseudo_const"() {value = dense<[[0.1]]> : tensor<1x1xf32>} : () -> tensor<1x1xf32>

  // Use a four dimension sharding (devices=[1,1,1,1]0)
  // Since the input tensor only has three dimensions, we expect this to fail.
  %0 = "tf.XlaSharding"(%arg0) { _XlaSharding = "\08\03\1A\04\01\01\01\01\22\01\00" } : (tensor<1x2x3xi32>) -> tensor<1x2x3xi32>
  %1 = "tf.A"(%0) : (tensor<1x2x3xi32>) -> (tensor<1x2x3xi32>)
  func.return %1: tensor<1x2x3xi32>
}

// -----

// CHECK-LABEL: func @check_retval_sharding_errors

  %output = "tf.Cast"(%q_output) {Truncate = false} : (tensor<1x2x2x1x!tf_type.qint32>) -> tensor<1x2x2x1xi32>

  // CHECK-DAG: %[[OUTPUT_2:.*]] = "tf.Cast"(%[[OUTPUT_QINT]]) <{Truncate = false}> : (tensor<1x2x2x1x!tf_type.qint32>) -> tensor<1x2x2x1xi32>
  // CHECK-DAG: %[[OUTPUT_QINT_1:.*]] = "tf.Cast"(%[[OUTPUT_1]]) <{Truncate = false}> : (tensor<1x2x2x1xi32>) -> tensor<1x2x2x1x!tf_type.qint32>

// CHECK-LABEL: func @testCase(%arg0: tensor<i32>, %arg1: tensor<!tf_type.resource<tensor<1x2x3xf32>>>)
func.func @testCase(%arg0: tensor<i32>, %arg1: tensor<!tf_type.resource<tensor<1x2x3xf32>>>) -> tensor<1x2x3xf32> {
  %0 = "tf.Case"(%arg0, %arg1) {branches = [@branch_0, @branch_1], is_stateless = false} : (tensor<i32>, tensor<!tf_type.resource<tensor<1x2x3xf32>>>) -> tensor<1x2x3xf32>

  // CHECK: [[Result0:%.*]] = "tf.CaseRegion"

  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";

  // CHECK-NEXT: %[[UPDATE:.*]] = "tf.XlaDynamicUpdateSlice"(%[[STACK_VAL]], %[[UPDATE_SLICE]], %[[CONCAT_OFFETS]]) : (tensor<10x2xi32>, tensor<1x2xi32>, tensor<2xi32>) -> tensor<10x2xi32>
  // CHECK-NEXT: "tf.AssignVariableOp"(%[[BUFFER]], %[[UPDATE]]) : (tensor<!tf_type.resource<tensor<10x2xi32>>>, tensor<10x2xi32>) -> ()
  // CHECK-NEXT: %[[CONST1:.*]] = "tf.Const"() <{value = dense<1> : tensor<1xi32>}> : () -> tensor<1xi32>