// function, which might not exist in case of multi-signature graphs. In that
// case, this pass will create a new main function, which calls signature
// functions.
//
// An already existing @main function will be renamed by attaching a numeric
// suffix like `@main_0` to avoid conflict with the newly created main function.
class InsertMainFunctionPass
    : public PassWrapper<InsertMainFunctionPass, OperationPass<ModuleOp>> {

                 Out(NodeWith(
                     Op("Slice"),
                     Name("begin/static_shaped_slice/static_shaped_slice"))),
                 _)));
}

// New constants being created need to have control dependencies copied to
// ensure correct control flow analysis in TF V2.
TEST(SliceToDynamicSliceRewriteTest, WithControlDepsToConstant) {
  Scope root = Scope::NewRootScope()

      builder.clone(*op, mapping);
    }
    // Creates the ReturnOp so that the per-host function returns the
    // correct values of the cloned operations.
    llvm::SmallVector<Value, 4> results_after_mapping;
    for (Value result : metadata.results) {
      results_after_mapping.push_back(mapping.lookupOrDefault(result));
    }
    builder.create<func::ReturnOp>(loc, results_after_mapping);

      if (pjrt_tensor_buffer != nullptr) {
        args->push_back(pjrt_tensor_buffer->pjrt_buffer());
      } else {
        // Creates a PjRtBuffer from DeviceMemoryBase. The newly created
        // PjRtBuffer needs to be persisted till XLA execution is completed.
        auto dmem = se::DeviceMemoryBase(
            const_cast<char*>(tensor->tensor_data().data()),

          !IsQintValueDefinedByIntToQintCast(op->getOperand(i))) {
        new_operands.push_back(rewriter.create<TF::CastOp>(
            op->getLoc(), orig_op_type, operands[i]));
      } else {
        new_operands.push_back(operands[i]);
      }
    }

    // Create a new UQ op.
    OperationState state(op->getLoc(), op->getName().getStringRef(),

  void MarkAsResult(AbstractTensorHandle* gradient) const override;

  void DeleteGradient(AbstractTensorHandle* gradient) const override;

 private:
  // The context where the aggregation op `Add` is to be created.
  AbstractContext* ctx_;
};

// Returns the number of elements in the gradient tensor.
int64_t TapeVSpace::NumElements(AbstractTensorHandle* tensor) const {

  llvm::SmallVector<mlir::TF::SplitOp, 4> split_ops_for_tiled_input;
  split_ops_for_tiled_input.reserve(
      input_sharding.tile_assignment_devices_size());

  // Creates a tree of split nodes for sharding tiled inputs. Splits nodes
  // are created such that input data is sharded in row major order.
  // Split nodes at ith depth from the original input node represent nodes
  // that split the input data at i-th dimension.

//     probe for gadgets inside core TensorFlow. We can even protect the area
//     of memory where the copies reside to not allow any more writes to it
//     after all copies are created.
template <typename T>
static std::unique_ptr<const T> CopyToCore(const T* plugin_ops,
                                           size_t plugin_size) {
  if (plugin_ops == nullptr) return nullptr;

  } else {
    return false;
  }
  return true;
}

// Create a dummy zero to be fed locally from the host to the TPUExecute.
Value CreateZeroInput(Location loc, OpBuilder& builder, Attribute zero_attr,
                      DenseIntElementsAttr shape_attr) {
  ConstOp zero = builder.create<ConstOp>(loc, zero_attr);
  zero->setAttr(kICIWeightDistributionMlirBridgeMarker,

  builder.setInsertionPoint(&terminator);
  builder.create<tf_executor::YieldOp>(terminator.getLoc(),
                                       terminator.getOperands());
  terminator.erase();

  builder.setInsertionPoint(island_op);
  IRMapping mapping;
  for (int i : llvm::seq<int>(0, num_replicas)) {
    // Create new island for replica.
    auto replica = builder.create<tf_executor::IslandOp>(