// CHECK: return %[[CST]]
  %0 = "tfl.transpose"(%cst, %cst_perm) : (tensor<2x3x4xi32>, tensor<3xi32>) -> tensor<4x2x3xi32>
  func.return %0 : tensor<4x2x3xi32>
}

// CHECK-LABEL: @ConstantFoldBinaryOpDynamicOutput
func.func @ConstantFoldBinaryOpDynamicOutput() -> tensor<?xi32> {

func.func @ConvertIdentityGatherNdOp3D(%arg0: tensor<4x3x4xf32>) -> tensor<4x3x4xf32> {
  %cst = arith.constant dense<[[0], [1], [2], [3]]> : tensor<4x1xi32>
  %0 = "tfl.gather_nd"(%arg0, %cst) : (tensor<4x3x4xf32>, tensor<4x1xi32>) -> tensor<4x3x4xf32>
  func.return %0 : tensor<4x3x4xf32>

// CHECK-LABEL: ConvertIdentityGatherNdOp3D
// CHECK-SAME: (%[[ARG:.*]]: tensor<4x3x4xf32>) -> tensor<4x3x4xf32>

func.func @convert_dot_general_dynamic_batch_dim(%arg0: tensor<2x?x2x3xf32>, %arg1: tensor<2x?x4x3xf32>) -> tensor<2x?x2x4xf32> {
%0 = "mhlo.dot_general"(%arg0, %arg1) {
  dot_dimension_numbers = #mhlo.dot<
    lhs_batching_dimensions = [0, 1],
    rhs_batching_dimensions = [0, 1],
    lhs_contracting_dimensions = [3],
    rhs_contracting_dimensions = [3]
  >} : (tensor<2x?x2x3xf32>, tensor<2x?x4x3xf32>) -> tensor<2x?x2x4xf32>

// CHECK:           %[[VAL_6:.*]] = arith.constant dense<[0, 3, 1, 2]> : tensor<4xi32>
// CHECK:           %[[VAL_7:.*]] = "tfl.transpose"(%[[VAL_5]], %[[VAL_6]]) : (tensor<1x4x4x3xf32>, tensor<4xi32>) -> tensor<1x3x4x4xf32>

@testLstmWithInvalidInputsRankMatch(%arg0: tensor<1x4xf32>, %arg1: tensor<4x4xf32>, %arg2: tensor<4x4xf32>, %arg3: tensor<4x4xf32>, %arg4: tensor<4x4xf32>, %arg5: tensor<4x4xf32>, %arg6: tensor<4x4xf32>, %arg7: tensor<4x4xf32>, %arg8: tensor<4x4xf32>, %arg9: tensor<4xf32>, %arg10: tensor<4xf32>, %arg11: tensor<4xf32>, %arg12: tensor<1x4xf32>, %arg13: tensor<4xf32>, %arg14: tensor<4xf32>, %arg15: tensor<4xf32>, %arg16: tensor<4x4xf32>, %arg17: tensor<4xf32>, %arg18: tensor<3xf32>, %arg19: tensor<3xf32>,...

  %3 = "tf.AddV2"(%arg0, %0): (tensor<4x4xf32>, tensor<1xf32>) -> tensor<4x4xf32>
  %4 = "tf.Log"(%3) {device = "/job:localhost/replica:0/task:0/device:GPU:0"}: (tensor<4x4xf32>) -> tensor<4x4xf32>

  // CHECK: %[[ADD1:.*]] = "tf.AddV2"
  // CHECK: %[[LOG1:.*]] = "tf.Log"(%[[ADD1]])
  %5 = "tf.AddV2"(%4, %1): (tensor<4x4xf32>, tensor<1xf32>) -> tensor<4x4xf32>

// %0:3 = "tf.SplitV"(%input, %split_sizes, %split_dim) :
//                   (tensor<4x6xf32>, tensor<3xi32>, tensor<i32>) ->
//                   (tensor<4x1xf32>, tensor<4x2xf32>, tensor<4x3xf32>)
//
// We will generate slices following slices:
// %0 = "mhlo.slice"(%input) {
//        limit_indices = dense<[4, 1]> : tensor<2xi64>,
//        start_indices = dense<0> : tensor<2xi64>,