std::vector<int64_t> targets;
  std::vector<Gradient*> used_in_grads;
  // We may end up with slightly fewer elements than we reserve, but grad.size()
  // should be a reasonably tight upper bound.
  targets.reserve(grad.size());
  used_in_grads.reserve(grad.size());
  std::unordered_map<int64_t, TapeTensor> sources_that_are_targets;
  for (int grad_index = 0, end = grad.size(); grad_index < end; ++grad_index) {

    ImmediateExecutionContext* ctx, std::unique_ptr<FlatTensorFunction>* out) {
  TF_RETURN_IF_ERROR(ctx->AddFunctionDef(*function_def));
  std::vector<ImmediateTensorHandlePtr> owned_captures;
  owned_captures.reserve(captures.size());
  for (ImmediateExecutionTensorHandle* capture : captures) {
    capture->Ref();
    owned_captures.push_back(ImmediateTensorHandlePtr(capture));
  }

  std::size_t dim_size = type.getDimSize(quantization_dim_);
  if (dim_size != scales_.size()) {
    return {};
  }
  SmallVector<UniformQuantizedValueConverter, 4> converters;
  converters.reserve(dim_size);
  for (int i = 0, e = dim_size; i != e; ++i) {
    converters.push_back(getPerChunkConverter(i));
  }

  // Scan the elements of the dense elements attributes and quantize them by

#include "xla/client/local_client.h"

namespace tensorflow {
namespace {
std::vector<const xla::Shape*> GetShapePointers(
    absl::Span<const xla::Shape> shapes) {
  std::vector<const xla::Shape*> shape_ptrs;
  shape_ptrs.reserve(shapes.size());
  for (const auto& shape : shapes) {
    shape_ptrs.push_back(&shape);
  }
  return shape_ptrs;
}
}  // namespace

absl::StatusOr<std::unique_ptr<xla::LocalExecutable>>

      // String
      if (const int s_size = default_value.list().s_size()) {
        absl::InlinedVector<const void*, 4> values_vector;
        values_vector.reserve(s_size);
        absl::InlinedVector<size_t, 4> lengths_vector;
        lengths_vector.reserve(s_size);
        for (int i = 0; i < s_size; ++i) {
          const string& v = default_value.list().s(i);
          values_vector.push_back(v.data());

    // Parameters for constructing each slice.
    SmallVector<Value, 4> begin_indices, end_indices, strides;
    begin_indices.reserve(value_rank);
    end_indices.reserve(value_rank);
    strides.reserve(value_rank);
    // final output shape
    SmallVector<Value, 4> shape_values;
    shape_values.reserve(value_rank - 1);
    // slice shape before reshape, should be like{?, 1, ?, ?} if axis = 1

    if (const_args[i]) {
      constant_arg_indices->push_back(i);
    }
  }

  // There can be hundreds of resource variables. Reserve the space for them.
  // We don't reserve for constants above as they are usually few.
  resource_arg_indices->reserve(arg_types.size());
  for (size_t i = 0; i < arg_types.size(); ++i) {
    if (arg_types[i] == DT_RESOURCE) {
      resource_arg_indices->push_back(i);
    }

  }
  return variable_lookup;
}

}  // anonymous namespace

std::vector<const Tensor*> InputsFromContext(OpKernelContext* ctx) {
  std::vector<const Tensor*> inputs;
  inputs.reserve(ctx->num_inputs());
  for (int input_idx = 0; input_idx < ctx->num_inputs(); input_idx++) {
    inputs.push_back(&ctx->input(input_idx));
  }
  return inputs;
}

  if (variable_names_set.empty()) {
    return OkStatus();
  }
  std::vector<string> variable_names;
  variable_names.reserve(variable_names_set.size());
  std::vector<string> tensor_names;
  tensor_names.reserve(variable_names_set.size());
  for (const string& node_name : variable_names_set) {
    variable_names.push_back(node_name);
    NodeDef* node_def = name_to_node_map.at(node_name);

	MOVW	$0, R0
	MRC	15, 0, R1, C13, C0, 2
	MOVW	R0, 0x34(R1)

	MOVW	8(R4), R12	// libcall->args

	// Do we have more than 4 arguments?
	MOVW	4(R4), R0	// libcall->n
	SUB.S	$4, R0, R2
	BLE	loadregs

	// Reserve stack space for remaining args
	SUB	R2<<2, R13
	BIC	$0x7, R13	// alignment for ABI

	// R0: count of arguments
	// R1:
	// R2: loop counter, from 0 to (n-4)
	// R3: scratch
	// R4: pointer to libcall struct