// CHECK-NEXT: [[res:%.*]] = tfrt_fallback_async.executeop key(1) cost({{.*}}) device("/job:localhost/replica:0/task:0/device:CPU:0") "tf.GatherV2"([[value]], {{.*}}, [[const]])
// CHECK-NEXT: tfrt.return [[out_chain]], [[res]] : !tfrt.chain, !tfrt_fallback.tf_tensor
func.func @gather(%indices: tensor<?xi32>,

// CHECK-NEXT: tf.Const
// CHECK-NEXT: tf.Const
// CHECK-NEXT: tf.Const
// CHECK-NEXT: tf.Const
// CHECK-NEXT: tf.GatherV2
// CHECK-NEXT: tf.AddV2
// CHECK-NEXT: tf.AddV2
// CHECK-NEXT: tf.GatherV2
// CHECK-NEXT: tf.Mul
// CHECK-NEXT: tf.GatherV2
// CHECK-NEXT: tf.Mul
// CHECK-NEXT: tf_mlrt.tf_await
// CHECK-NEXT: tf_mlrt.tf_await
// CHECK-NEXT: tf.TensorListSetItem

  }
}


// -----
//  Decomposable Resource Ops usage
//
// CHECK-LABEL:  func.func @serving_default
// CHECK:      "tf.VarHandleOp"
// CHECK-NEXT: "tf.IfrtLoadVariable"
// CHECK-NEXT: "tf.GatherV2"
// CHECK-NEXT:  return 
//
module {
  func.func @serving_default() -> tensor<1x3xbf16> {
    %cst = "tf.Const"() <{value = dense<[1]> : tensor<1xi32>}> : () -> tensor<1xi32>

  %383 = "tf.GatherV2"(%11, %382#1, %14) {batch_dims = 0 : i64, device = ""} : (tensor<1xi64>, tensor<?xi64>, tensor<i32>) -> tensor<?xi64>
  %384 = "tf.GatherV2"(%12, %383, %14) {batch_dims = 0 : i64, device = ""} : (tensor<1xi64>, tensor<?xi64>, tensor<i32>) -> tensor<?xi64>

  if (mlir::isa<mlir::TF::StringType>(key_type.getElementType()))
    cost *= kLookupTableFindStringKeyCostScale;

  return cost;
}

// The cost function for tf.GatherV2.
int64_t InferGatherV2Cost(const CostContext& context, mlir::TF::GatherV2Op op) {
  return InferTensorSize(
      context, mlir::cast<mlir::TensorType>(op.getOutput().getType()));
}

  std::set<string> constrained_kernel = {"Diag",
                                         "DiagPart",
                                         "GatherNd",
                                         "GatherV2",
                                         "QuantizeAndDequantizeV2",
                                         "QuantizeAndDequantizeV3",
                                         "QuantizeAndDequantizeV4",

  }
  func.func private @composite_gather_fn_1(%arg0: tensor<1024x3x4x3xf32>, %arg1: tensor<1xi32>, %arg2: tensor<i32>) -> tensor<1x3x4x3xf32> attributes {tf_quant.composite_function} {
    %0 = "tf.GatherV2"(%arg0, %arg1, %arg2) {attr_map = "0:batch_dims", batch_dims = 0 : i64, device = ""} : (tensor<1024x3x4x3xf32>, tensor<1xi32>, tensor<i32>) -> tensor<1x3x4x3xf32>
    return %0 : tensor<1x3x4x3xf32>
  }

                         %out_scale : tensor<*xf32>, %out_zp : tensor<*xi32>) -> tensor<*xi8>
      attributes {tf_quant.quantized_ops = ["Gather"]} {

    %out = "tf.GatherV2"(%weight, %input, %axis) {
      batch_dims = 0 : i64, attr_map = "batch_dims:0"} : (tensor<*xi8>, tensor<*xi32>, tensor<i32>) -> tensor<*xi8>

      (TF_AssignVariableOp $v_resource, $new_v, (CreateConstBoolAttrFalse))
    ]
  >;

// Pattern to decompose tf.ResourceGather into tf.ReadVariable and tf.GatherV2.
def DecomposeResourceGather : Pat<
  (TF_ResourceGatherOp:$old_result
    $resource, $indices, $batch_dims, $validate_indices),
  (TF_GatherV2Op:$dest
    (CreateTFReadVariableOp $old_result, $old_result, $resource),

    %resource = "tf.VarHandleOp"() {container = "c", shared_name = "v"} : () -> tensor<*x!tf_type.resource<tensor<*xi32>>>

    // CHECK-DAG: [[READVAR:%.+]] = "tf.ReadVariableOp"([[VAR]])
    // CHECK: [[GATHER:%.+]] = "tf.GatherV2"([[READVAR]], [[INDEX]], [[ZERO]]) <{batch_dims = 0 : i64}> {_xla_outside_compilation = "0"} : (tensor<*xi32>, tensor<?xi32>, tensor<i64>) -> tensor<*xi32>
    // CHECK: return [[GATHER]]