%4 = stablehlo.convolution(%arg0, %0) dim_numbers = [b, f, 0, 1]x[o, i, 0, 1]->[b, f, 0, 1], window = {pad = [[1, 1], [1, 1]]} {batch_group_count = 1 : i64, feature_group_count = 1 : i64} : (tensor<1x2x5x5xf32>, tensor<4x2x3x3xf32>) -> tensor<1x4x5x5xf32>
  %5 = stablehlo.add %4, %3 : tensor<1x4x5x5xf32>
  %6 = stablehlo.maximum %5, %2 : tensor<1x4x5x5xf32>
  return %6 : tensor<1x4x5x5xf32>
}

func.func @move_across_double_transpose(%arg0: tensor<1x4x4x8xf32>, %arg1: tensor<1x4x4x8xf32>) -> tensor<1x4x8x4xf32> {

  // CHECK: %[[RES_PERM:.*]] = "tf.Const"() <{value = dense<[0, 3, 1, 2]> : tensor<4xi32>}>
  // CHECK: %[[ADD:[0-9]*]] = "tf.AddV2"(%arg0, %arg1) : (tensor<1x4x4x8xf32>, tensor<1x4x4x8xf32>) -> tensor<1x4x4x8xf32>
  // CHECK: %[[RES_TRANSPOSE_0:[0-9]*]] = "tf.Transpose"(%[[ADD]], %[[RES_PERM]])

// CHECK: return %6 : tensor<1x5x5x3xf32>
}

// CHECK-LABEL: bias_adjust_duplicate_filter
func.func @bias_adjust_duplicate_filter(%arg0: tensor<1x5x5x2xf32>) -> (tensor<1x5x5x3xf32>, tensor<1x5x5x3xf32>) {
  %0 = "quantfork.stats"(%arg0) {
    layerStats = dense<[-1.28e-5, 1.27e-5]> : tensor<2xf32>
  } : (tensor<1x5x5x2xf32>) -> tensor<1x5x5x2xf32>

  %0 = stablehlo.constant dense<2.000000e+00> : tensor<1x4x3x3xf32>
  %1 = stablehlo.transpose %arg0, dims = [0, 3, 1, 2] : (tensor<1x3x3x4xf32>) -> tensor<1x4x3x3xf32>
  %2 = stablehlo.add %1, %0 : tensor<1x4x3x3xf32>
  return %2 : tensor<1x4x3x3xf32>
}
// CHECK-SAME: (%[[ARG_0:.+]]: tensor<1x3x3x4xf32>) -> tensor<1x4x3x3xf32>
// CHECK-DAG: %[[CONST_0:.+]] = stablehlo.constant

  %0:2 = "tf.SplitV"(%arg0, %arg1, %arg2) : (tensor<1x4x3x3xf32>, tensor<2xi32>, tensor<i32>) -> (tensor<1x4x2x3xf32>, tensor<1x4x1x3xf32>)
  func.return %0#0 : tensor<1x4x2x3xf32>

  // CHECK-LABEL: splitv
  // CHECK: "tfl.split_v"(%arg0, %arg1, %arg2) <{num_splits = 2 : i32}> : (tensor<1x4x3x3xf32>, tensor<2xi32>, tensor<i32>) -> (tensor<1x4x2x3xf32>, tensor<1x4x1x3xf32>)
}

  return %0 : tensor<1x1x1x5xf32>
}
func.func private @XlaCallModule_aten.avg_pool2d.default.impl_6(%arg0: tensor<1x1x1x8xf32>) -> tensor<1x1x1x5xf32>

// CHECK-LABEL: avg_pool2d_7
// CHECK: %cst = arith.constant dense<[0, 2, 3, 1]> : tensor<4xi32>

  %1 = "tf.AddV2"(%0, %0) : (tensor<1x4x4x8xf32>, tensor<1x4x4x8xf32>) -> tensor<1x4x4x8xf32>
  %2 = "tf.Const"() {value = dense<[0, 3, 1, 2]> : tensor<4xi32>} : () -> tensor<4xi32>
  %3 = "tf.Transpose"(%1, %2) : (tensor<1x4x4x8xf32>, tensor<4xi32>) -> tensor<1x8x4x4xf32>

  func.return %3 : tensor<1x8x4x4xf32>
}

// CHECK-LABEL: move_transpose_handle_broadcast

// CHECK: %[[CUSTOM_AGGREGATOR_1:.+]], {{.*}}, {{.*}}, {{.*}} = "tf.CustomAggregator"(%[[XLA_CALL_MODULE]]) {{.*}} : (tensor<1x4x4x8xf32>) -> (tensor<1x4x4x8xf32>, tensor<f32>, tensor<f32>, tensor<0xi64>)

// CHECK: %[[TRANSPOSE_2:.+]] = stablehlo.transpose %[[CUSTOM_AGGREGATOR_1]], dims = [0, 3, 1, 2] : (tensor<1x4x4x8xf32>) -> tensor<1x8x4x4xf32>

      %1 = "tf.XlaCallModule"(%0, %cst_0, %cst_1) <{Sout = [#tf_type.shape<1x4x4x3>], module = "", version = 9 : i64}> {_entry_function = @composite_fn_2, _original_entry_function = "composite_fn_2", _tfl_quant_trait = "fully_quantizable"} : (tensor<1x4x4x3xf32>, tensor<3x1x1x3xf32>, tensor<3xf32>) -> tensor<1x4x4x3xf32>
      return %1 : tensor<1x4x4x3xf32>

// RUN: tf-opt %s -split-input-file -verify-diagnostics

func.func @out_of_bounds_check(%arg0: tensor<1x4x4x8xf32>) -> tensor<1x4x4x8xf32> {
  %0 = "tf.Const"() {value = dense<[0, 3, 1, 2]> : tensor<4xi32>} : () -> tensor<4xi32>
  %1 = "tf.Const"() {value = dense<[0, 0x4141, 3, 1]> : tensor<4xi32>} : () -> tensor<4xi32>
  %2 = "tf.Transpose"(%arg0, %0) : (tensor<1x4x4x8xf32>, tensor<4xi32>) -> tensor<1x8x4x4xf32>