start := b.pb.startMessage()
	b.pb.uint64Opt(tagMapping_ID, id)
	b.pb.uint64Opt(tagMapping_Start, base)
	b.pb.uint64Opt(tagMapping_Limit, limit)
	b.pb.uint64Opt(tagMapping_Offset, offset)
	b.pb.int64Opt(tagMapping_Filename, b.stringIndex(file))
	b.pb.int64Opt(tagMapping_BuildID, b.stringIndex(buildID))

	// Output profile in protobuf form.
	b := newProfileBuilder(w)
	b.pbValueType(tagProfile_PeriodType, countName, "count")
	b.pb.int64Opt(tagProfile_Period, 1)
	b.pbValueType(tagProfile_SampleType, countName, "count")
	b.pbValueType(tagProfile_SampleType, cycleName, "nanoseconds")

	cpuGHz := float64(pprof_cyclesPerSecond()) / 1e9

namespace odml {
namespace {
// A struct to hold axes and sizes for a set of dimensions.
struct DimensionVector {
  llvm::ArrayRef<int64_t> AxesArray() const { return axes; }
  llvm::ArrayRef<int64_t> SizesArray() const { return sizes; }

  llvm::SmallVector<int64_t, 4> axes;
  llvm::SmallVector<int64_t, 4> sizes;
};

// Appends all elements in `range` to `values`.
template <typename ValueT, typename Range>

namespace mlir {
namespace odml {

bool IsIotaAttr(ArrayRef<int64_t> arr, int64_t size) {
  if (arr.size() != size) return false;
  int64_t iota = 0;
  for (auto s : arr) {
    if (s != iota) return false;
    ++iota;
  }
  return true;
}

PermutationAndShape GetPermutationAndTransposedShape(
    llvm::ArrayRef<int64_t> permutation_array, ShapedType input_type,
    ConversionPatternRewriter& rewriter) {

                               ArrayRef<int64_t> index) {
#ifndef NDEBUG
    assert(shape.size() == index.size());
    for (size_t i = 0; i < shape.size(); ++i) {
      assert(index[i] < shape[i]);
      assert(index[i] >= 0);
    }
#endif  // NDEBUG
    int64_t offset = 0;
    int64_t stride = 1;
    for (int64_t dim = shape.size() - 1; dim >= 0; --dim) {

  // but not in the LHS.
  std::vector<int64_t> lhs;
  std::vector<int64_t> rhs;
  std::vector<std::tuple<int64_t, int64_t>> lhs_rhs;
  std::vector<std::tuple<int64_t, int64_t>> lhs_out;
  std::vector<std::tuple<int64_t, int64_t>> rhs_out;
  std::vector<std::tuple<int64_t, int64_t, int64_t>> lhs_rhs_out;
};

TF::ReshapeOp createOutputReshapeOpForDynamic(
    Value value, ArrayRef<int64_t> shape, Value org_lhs, Value org_rhs,

}

Value CreateStridedSliceOp(mlir::Location loc, ArrayRef<int64_t> output_shape,
                           Value input, ArrayRef<int32_t> begin,
                           ArrayRef<int32_t> end, ArrayRef<int32_t> strides,
                           int64_t begin_mask, int64_t end_mask,
                           int64_t ellipsis_mask, int64_t new_axis_mask,
                           int64_t shrink_axis_mask, OpBuilder* builder) {

// calculate padding_low and padding_high for SAME padding.
void GetSamePaddingValues(OpBuilder &builder, Location loc, Value input_size,
                          int64_t filter_sz, int64_t dilation_rate,
                          int64_t stride, Value &padding_low,
                          Value &padding_high) {
  Value zero = CreateScalarConstValue<int32_t>(builder, loc, 0);

  TapeVSpace vspace(ctx);
  std::vector<int64_t> target_tensor_ids = MakeTensorIDList(targets);
  std::vector<int64_t> source_tensor_ids = MakeTensorIDList(sources);
  tensorflow::gtl::FlatSet<int64_t> sources_set(source_tensor_ids.begin(),
                                                source_tensor_ids.end());
  std::unordered_map<int64_t, TapeTensor> sources_that_are_targets;

  // Map from tensor ID to how many references still exist for this tensor in
  // the tape.
  std::unordered_map<int64_t, int64_t> tensor_usage_counts;

  // Maps from op ID to how many output tensors of this op still need to have
  // their gradients computed.
  std::unordered_map<int64_t, int64_t> op_missing_tensor;
};

// If `persistent_tape` is true, op_tape is not changed and none of the