}

  auto c = op.input(0);
  auto x = op.input(1);
  auto y = op.input(2);

  auto zeros = ZerosLike(scope, grad_inputs[0]);
  auto gx = SelectV2(scope, c, grad_inputs[0], zeros);
  auto x_shape = Shape(scope, x);
  auto output_shape = Shape(scope, op.output(0));

  // Reduce away broadcasted leading dims.

TEST(DeviceCompilationCacheTest, StoreMultipleEntries) {
  auto cache = std::make_unique<Cache>();

  TF_ASSERT_OK_AND_ASSIGN(auto key1, BuildSampleSignature("foo"));
  TF_ASSERT_OK_AND_ASSIGN(auto key2, BuildSampleSignature("bar"));

  auto compilation_result1 = std::make_unique<XlaCompiler::CompilationResult>();
  auto compilation_result2 = std::make_unique<XlaCompiler::CompilationResult>();
  auto executable1 = std::make_unique<FakeExecutable>("foo_exe");

}

void Exporter::UseOriginalFunctionNames(NodeDef& node_def) {
  if (!configs_.export_original_tf_func_name) return;

  auto& attrs = *node_def.mutable_attr();

  auto try_use_original_func_name = [this](std::string* name) {
    if (auto func = symbol_table_.lookup<FuncOp>(*name)) {
      if (auto original_func_name =
              func->getAttrOfType<mlir::StringAttr>("tf._original_func_name")) {

}

void Exporter::UseOriginalFunctionNames(NodeDef& node_def) {
  if (!configs_.export_original_tf_func_name) return;

  auto& attrs = *node_def.mutable_attr();

  auto try_use_original_func_name = [this](std::string* name) {
    if (auto func = symbol_table_.lookup<FuncOp>(*name)) {
      if (auto original_func_name =
              func->getAttrOfType<mlir::StringAttr>("tf._original_func_name")) {

  for (auto& input : inputs) {
    auto input_op = input.get().getDefiningOp();
    if (!input_op || !IsSupportedHostInputOp(input_op)) return false;
  }
  for (auto entry : llvm::enumerate(inputs)) {
    Value input = entry.value().get();
    auto ranked_type = mlir::dyn_cast<RankedTensorType>(input.getType());
    if (!ranked_type) return false;
    auto input_shape = ranked_type.getShape();

    return success();
  }

  auto stride_h = op.getStrideHAttr().getInt();
  auto stride_w = op.getStrideWAttr().getInt();
  auto dilation_h = op.getDilationHFactorAttr().getInt();
  auto dilation_w = op.getDilationWFactorAttr().getInt();

  // We don't have EXPLICIT PADDING in TfLite.
  auto paddings = op.getPadding();
  tensorflow::Padding padding;

  MLIRContext context;
  auto id =
      StringAttr::get(&context, "//tensorflow/python/keras/keras_file.py");
  auto loc = FileLineColLoc::get(&context, id, 0, 0);
  auto id2 =
      StringAttr::get(&context, "//tensorflow/python/something/my_test.py");
  auto loc2 = FileLineColLoc::get(&context, id2, 0, 0);
  auto id3 = StringAttr::get(&context, "python/tensorflow/show_file.py");
  auto loc3 = FileLineColLoc::get(&context, id3, 0, 0);

    return rewriter.notifyMatchFailure(op, "zero_points must be constant");
  }

  auto original_element_type = getElementTypeOrSelf(original_type);
  if (!mlir::isa<TF::Qint8Type, TF::Qint32Type>(original_element_type)) {
    return rewriter.notifyMatchFailure(
        op, "Quantized type must be qint8 or qint32.");
  }
  auto storage_type = GetIntTypeFromTFQint(original_element_type);

  llvm::SmallPtrSet<Operation*, 4> defining_ops;
  llvm::SmallVector<Type, 4> types;
  for (auto& item :
       llvm::make_early_inc_range(llvm::reverse(graph_op.GetBody()))) {
    auto it = new_control_inputs.find(&item);
    if (it == new_control_inputs.end()) continue;
    auto& new_control_inputs_for_item = it->second;
    builder.setInsertionPoint(&item);

                manifest {
                    attributes 'Automatic-Module-Name': 'auto'
                }
            }

        """
        file('auto-module/src/main/java/auto/AutoClass.java')  << """
            package auto;

            public class AutoClass {
                public void m() {
                    // can access, because a auto module sees the unnamed module
                    new moda.ModaClass();