func.func @foo(%arg0: tensor<1x4xf32>, %arg1: tensor<1x4xf32>) -> tensor<1x4xf32> {
  func.return %arg0 : tensor<1x4xf32>
}

func.func @bar(%arg0: tensor<2x4xf32>, %arg1: tensor<2x4xf32>) -> tensor<2x4xf32> {
  func.return %arg0 : tensor<2x4xf32>
}

func.func @foobar(%arg0: tensor<1x4xf32>, %arg1: tensor<1x4xf32>) -> tensor<1x4xf32> {
  func.return %arg0 : tensor<1x4xf32>
}

// CHECK-DAG: func @main

  %ch0 = tfrt.new.chain

  %a1 = tfrt_fallback_async.const_dense_tensor dense<[[2, 2], [2, 2]]> : tensor<2x2xi32>
  %a2 = tfrt_fallback_async.const_dense_tensor dense<[[3, 3], [3, 3]]> : tensor<2x2xi32>
  %b = tfrt_fallback_async.const_dense_tensor dense<[[1, 1], [1, 1]]> : tensor<2x2xi32>

  // Two batch_size=2 batches get concatenated.

  %cst = arith.constant dense<[2, 2]> : tensor<2xi32>
  %cst_0 = arith.constant dense<4> : tensor<2x2xi32>
  %cst_1 = arith.constant dense<0> : tensor<2x2xi32>
  %cst_2 = arith.constant dense<0> : tensor<4x2xi32>
  %0 = "tf.SpaceToBatchND"(%arg0, %cst, %cst_0) : (tensor<1x128x128x3xf32>, tensor<2xi32>, tensor<2x2xi32>) -> tensor<4x68x68x3xf32>

        tensor<1x1x5xf32>,
        tensor<2x5xf32>, tensor<2x5xf32>, tensor<2x5xf32>, tensor<2x5xf32>,
        tensor<2x4xf32>, tensor<2x4xf32>, tensor<2x4xf32>, tensor<2x4xf32>,
        tensor<2xf32>, tensor<2xf32>, tensor<2xf32>,
        tensor<2xf32>, tensor<2xf32>, tensor<2xf32>, tensor<2xf32>,
        tensor<4x2xf32>, tensor<4xf32>,
        tensor<1x4xf32>, tensor<1x2xf32>,
        none, none, none, none) -> tensor<*xf32>

  %group_assignment = "tf.Const"() { value = dense<[[0],[1]]> : tensor<2x1xi32> } : () -> tensor<2x1xi32>
  // CHECK: "mhlo.all_reduce"
  // CHECK{LITERAL}: replica_groups = dense<[[0], [1]]> : tensor<2x1xi64>
  // CHECK-NOT: channel_handle
  // CHECK: mhlo.add
  %0 = "tf.XlaAllReduce"(%input, %group_assignment) {reduce_op = "Add", mode = "CrossReplica"} : (tensor<f32>, tensor<2x1xi32>) -> tensor<f32>
  func.return %0 : tensor<f32>
}

// -----

// CHECK: [[UNPACK:%.*]]:3 = "tf.Unpack"(%arg0) <{axis = 0 : i64}> : (tensor<3x2x2xf32>) -> (tensor<2x2xf32>, tensor<2x2xf32>, tensor<2x2xf32>)
// CHECK: [[SCALAR_ZERO:%.*]] = arith.constant dense<0> : tensor<i32>
// CHECK: [[CONCAT:%.*]] = "tf.Concat"([[SCALAR_ZERO]], [[UNPACK]]#0, [[UNPACK]]#1, [[UNPACK]]#2) : (tensor<i32>, tensor<2x2xf32>, tensor<2x2xf32>, tensor<2x2xf32>) -> tensor<?x2xf32>

    %0 = "tf.VarHandleOp"() {container = "c", shared_name = "v"} : () -> tensor<*x!tf_type.resource<tensor<2x8xi32>>>
    %1 = "tf.VarHandleOp"() {container = "c", shared_name = "w"} : () -> tensor<*x!tf_type.resource<tensor<2xi32>>>
    %2 = "tf.Const"() {value = dense<1> : tensor<i32>} : () -> tensor<i32>
    %3 = "tf.Const"() {value = dense<1> : tensor<i32>} : () -> tensor<i32>

  %4 = "tf.MatMul"(%arg0, %3) {device = "", transpose_a = false, transpose_b = false} : (tensor<2x3xf32>, tensor<3x4xf32>) -> tensor<2x4xf32>
  %5 = "tf.Identity"(%4) {device = ""} : (tensor<2x4xf32>) -> tensor<2x4xf32>
  %6 = "tf.Identity"(%5) {device = ""} : (tensor<2x4xf32>) -> tensor<2x4xf32>
  func.return %6 : tensor<2x4xf32>

module attributes {tf.versions = {producer = 888 : i32}} {
func.func @testReshapeShapeInference(%arg0: tensor<3x4xi32>) -> tensor<*xi32> {
  %cst = arith.constant dense<[1, 6, 2]> : tensor<3xi32>
  // CHECK: "tfl.reshape"(%arg0, %cst) : (tensor<3x4xi32>, tensor<3xi32>) -> tensor<1x6x2xi32>
  %0 = "tfl.reshape"(%arg0, %cst) : (tensor<3x4xi32>, tensor<3xi32>) -> tensor<*xi32>
  func.return %0 : tensor<*xi32>
}
}

// -----

  %3 = "tfl.dequantize"(%2) : (tensor<2x3x!quant.uniform<i16:f32, 1.0>>) -> (tensor<2x3xf32>)
  %4 = "tfl.concatenation"(%1, %3) {axis = -1 : i32, fused_activation_function = "NONE"} : (tensor<2x1xf32>, tensor<2x3xf32>) -> tensor<2x4xf32>
  %5 = "tfl.add"(%4, %arg2) {fused_activation_function = "NONE"} : (tensor<2x4xf32>, tensor<2x4xf32>) -> tensor<2x4xf32>
  func.return %5: tensor<2x4xf32>