%4 = "tf.LayerNorm"(%a1, %a2, %a3, %a4) {_tfl_quant_trait = "fully_quantizable", device = ""} : (tensor<128x128xf32>, tensor<1xf32>, tensor<1xf32>, tensor<1xi32>) -> tensor<128x128xf32>
   "tfl.yield"(%4) : (tensor<128x128xf32>) -> ()
  }) {_tfl_quant_trait = "fully_quantizable", device = ""} : (tensor<128x128xf32>, tensor<1xf32>, tensor<1xf32>, tensor<1xi32>) -> tensor<128x128xf32>

  return %0 : tensor<*xi32>
}
func.func private @_batched(%arg0: tensor<1xi32>) -> tensor<1xi32> {
  return %arg0 : tensor<1xi32>
}

}

// -----

module attributes {tf_saved_model.semantics} {

// Test hoisting write side-effect ops.

  func.return %0 : tensor<4xi16>
}

func.func @i32() -> tensor<4xi32> {
  // CHECK-LABEL: @i32
  // CHECK: value = dense<[1, 2, 3, 16909060]> : tensor<4xi32>
  // Check bytes come back in the right order
  %0 = "tfl.pseudo_const" () { value = dense<[1, 2, 3, 16909060]> : tensor<4xi32> } : () -> tensor<4xi32>
  func.return %0 : tensor<4xi32>
}

func.func @i64() -> tensor<4xi64> {
  // CHECK-LABEL: @i64

    %0 = "tf.Cast"(%input) {Truncate = false} : (tensor<*xi8>) -> tensor<*xi32>
    %1 = "tf.Sub"(%0, %input_zp) : (tensor<*xi32>, tensor<*xi32>) -> tensor<*xi32>

    // Use identity op to avoid the weight being constant-folded.
    %identity = "tf.Identity"(%filter) : (tensor<*xi8>) -> tensor<*xi8>
    %2 = "tf.Cast"(%identity) {Truncate = false} : (tensor<*xi8>) -> tensor<*xi32>

// CHECK-NEXT: %[[SHAPE:.*]] = "tf.RiscShape"(%arg0) {T = i32} : (tensor<2x3xf32>) -> tensor<*xi32>
// CHECK-NEXT: %[[ALPHA1:.*]] = "tf.RiscBroadcast"(%[[ALPHA]], %[[SHAPE]]) : (tensor<f32>, tensor<*xi32>) -> tensor<*xf32>
// CHECK-NEXT: %[[MAX:.*]] = "tf.RiscMaximum"(%arg0, %[[ALPHA1]]) : (tensor<2x3xf32>, tensor<*xf32>) -> tensor<*xf32>

    func.func @main(%arg0: tensor<3x3x10xbf16>, %arg1: tensor<3xi32>) -> tensor<1x?x4xbf16> attributes {allow_soft_placement = false, tf.entry_function = {control_outputs = "", inputs = "_arg0,_arg1,_arg2", outputs = "_retval0"}} {
      %cst = "tf.Const"() {value = dense<[1, -1, 4]> : tensor<3xi32>} : () -> tensor<3xi32>

  func.return %0 : tensor<4xi16>
}

func.func @i32() -> tensor<4xi32> {
  // CHECK-LABEL: @i32
  // CHECK: value = dense<[1, 2, 3, 16909060]> : tensor<4xi32>
  // Check bytes come back in the right order
  %0 = "tfl.pseudo_const" () { value = dense<[1, 2, 3, 16909060]> : tensor<4xi32> } : () -> tensor<4xi32>
  func.return %0 : tensor<4xi32>
}

func.func @i64() -> tensor<4xi64> {
  // CHECK-LABEL: @i64

  %2 = "tfl.pseudo_const"() {value = dense<[-1, 3]> : tensor<2xi32>} : () -> tensor<2xi32>
  %3 = "tfl.reshape"(%1, %2) : (tensor<1x1x3xf32>, tensor<2xi32>) -> tensor<1x3xf32>
  %4 = "quantfork.stats"(%3) {layerStats = dense<[-1.0, 1.0]> : tensor<2xf32>} : (tensor<1x3xf32>) -> tensor<1x3xf32>
  func.return %4 : tensor<1x3xf32>

// CHECK: %[[cst:.*]] = arith.constant dense<[-1, 3]> : tensor<2xi32>

  // CHECK-DAG: %[[X_SHAPE:.+]] = shape.shape_of %[[X]] : tensor<3x4x256x6xf32> -> tensor<4xindex>
  // CHECK-DAG: %[[EPS:.+]] = mhlo.constant dense<1.000000e+00> : tensor<256xf32>
  // CHECK-DAG: %[[MEAN:.+]] = "tf.Mean"(%arg0, %[[CST_AXIS]]) <{keep_dims = false}> : (tensor<3x4x256x6xf32>, tensor<3xi32>) -> tensor<256xf32>

//          (tensor<?x3x4x2xf32>) -> tensor<i32>
// %5 = stablehlo.reshape %4 :
//          (tensor<i32>) -> tensor<1xi32>
// %6 = stablehlo.concatenate %5, %0, %1, %2, dim = 0 :
//          (tensor<1xi32>, tensor<1xi32>, tensor<1xi32>, tensor<1xi32>)
//            -> tensor<4xi32>
// %7 = stablehlo.dynamic_broadcast_in_dim %arg2, %6
// %8 = stablehlo.add %3, %7
// ```
//