func.return %0 : tensor<4xui8>
}

func.func @int4() -> tensor<5xi4> {
  // CHECK-LABEL: @int4
  // CHECK: value = dense<[6, 3, -2, -7, -8]> : tensor<5xi4>
  %0 = "tfl.pseudo_const"() {value = dense<[6, 3, -2, -7, -8]> : tensor<5xi4>} : () -> tensor<5xi4>
  func.return %0 : tensor<5xi4>
}

func.func @qi32_per_axis() -> tensor<3x3x!quant.uniform<i32:f32:1, {1.0, 0.5:1, 0.25:1}>> {

  func.return %0 : tensor<4xui8>
}

func.func @int4() -> tensor<5xi4> {
  // CHECK-LABEL: @int4
  // CHECK: value = dense<[6, 3, -2, -7, -8]> : tensor<5xi4>
  %0 = "tfl.pseudo_const"() {value = dense<[6, 3, -2, -7, -8]> : tensor<5xi4>} : () -> tensor<5xi4>
  func.return %0 : tensor<5xi4>
}

func.func @qi32_per_axis() -> tensor<3x3x!quant.uniform<i32:f32:1, {1.0, 0.5:1, 0.25:1}>> {

  // be packed low-bits-first as [0x38, 0xBE] or [56, 190]. Tensor type should
  // be INT4.
  %0 = "tfl.pseudo_const" () {value = dense<[-8, 3, -2, -5]> : tensor<4xi4>} : () -> tensor<4xi4> loc("Const")
  func.return %0 : tensor<4xi4>

// -----

// CHECK-LABEL: @if
func.func @if() {
  // CHECK-NOT: Optional
  %0 = builtin.unrealized_conversion_cast to tensor<5xi1>
  %1 = "tf.OptionalNone"() : () -> tensor<!tf_type.variant<tensor<f32>>>
  %2 = "tf.If"(%0, %1) <{else_branch = @false, is_stateless = false, then_branch = @true}>
      : (tensor<5xi1>, tensor<!tf_type.variant<tensor<f32>>>) -> (tensor<!tf_type.variant<tensor<f32>>>)
  return
}

  %98 = "tf.Reshape"(%97, %cst_2) : (tensor<7x1x8x6x1x6x1x1x18xbf16>, tensor<5xi64>) -> tensor<7x8x6x6x18xbf16>
  func.return %98 : tensor<7x8x6x6x18xbf16>

  // CHECK-DAG: %[[CST:.*]] = "tf.Const"() <{value = dense<[1, 8, 1, 1, 18]> : tensor<5xi64>}> : () -> tensor<5xi64>
  // CHECK-DAG: %[[CST1:.*]] =  "tf.Const"() <{value = dense<[7, 8, 6, 6, 18]> : tensor<5xi64>}> : () -> tensor<5xi64>

  %1 = "tfl.shape"(%conv_s) : (tensor<?x26x26x26x16xf32>) -> tensor<5xi64>
  %2 = "tfl.broadcast_args"(%1, %cst) : (tensor<5xi64>, tensor<1xi64>) -> tensor<5xi64>
  %broad1 = "tfl.broadcast_to"(%conv_s, %2) : (tensor<?x26x26x26x16xf32>, tensor<5xi64>) -> tensor<?x26x26x26x16xf32>
  %broad2 = "tfl.broadcast_to"(%b, %2) : (tensor<16xf32>, tensor<5xi64>) -> tensor<?x26x26x26x16xf32>

func.func @testI8ToI4WithinRange() -> (tensor<4xi8>) {
  %0 = arith.constant dense<[-8, 0, 1, 7]> : tensor<4xi8>
  // expected-error@+1 {{type of return operand 0 ('tensor<4xi4>') doesn't match function result type ('tensor<4xi8>')}}
  func.return %0 : tensor<4xi8>
}

func.func @testI8ToI4NotWithinRange() -> tensor<4xi8> {
  %0 = arith.constant dense<[-10, 2, 3, 8]> : tensor<4xi8>

}

// CHECK-LABEL: func @testEmpty64
func.func @testEmpty64() -> (tensor<5xi64>) {
  %0 = "tf.Const"() { value = dense<5> : tensor<i32> } : () -> tensor<i32>

  // CHECK: [[VAL:%.+]] = "tf.Const"() <{value = dense<0> : tensor<5xi64>}>
  // CHECK: return [[VAL]] : tensor<5xi64>
  %1 = "tf.Empty"(%0) : (tensor<i32>) -> (tensor<5xi64>)
  func.return %1 : tensor<5xi64>
}

// CHECK-LABEL: func @testEmptyFloat

      "mhlo.return"(%5) : (tensor<f32>) -> ()
    }) {
    base_dilations = dense<1> : tensor<5xi64>,
    padding = dense<0> : tensor<5x2xi64>,
    window_dilations = dense<1> : tensor<5xi64>,
    window_dimensions = dense<[1, 3, 3, 3, 1]> : tensor<5xi64>,
    window_strides = dense<[1, 2, 2, 2, 1]> : tensor<5xi64>} : (tensor<4x16x16x16x8xf32>, tensor<f32>) -> tensor<4x7x7x7x8xf32>
  %3 = mhlo.divide %2, %1 : tensor<4x7x7x7x8xf32>

    %cst = arith.constant dense<[2, 1, 200, 200, 200]> : tensor<5xi64>
    %0 = "tfl.concatenation"(%arg0, %arg1) {axis = 0 : i32, fused_activation_function = "NONE"} : (tensor<1x200x200x200xf32>, tensor<1x200x200x200xf32>) -> tensor<2x200x200x200xf32>
    %1 = "tfl.reshape"(%0, %cst) : (tensor<2x200x200x200xf32>, tensor<5xi64>) -> tensor<2x1x200x200x200xf32>
    func.return %1 : tensor<2x1x200x200x200xf32>
  }