docker pull "$TFCI_DOCKER_IMAGE" || sleep 15
  docker pull "$TFCI_DOCKER_IMAGE"
fi 

if [[ "$TFCI_DOCKER_REBUILD_ENABLE" == 1 ]]; then
  DOCKER_BUILDKIT=1 docker build --cache-from "$TFCI_DOCKER_IMAGE" -t "$TFCI_DOCKER_IMAGE" $TFCI_DOCKER_REBUILD_ARGS
  if [[ "$TFCI_DOCKER_REBUILD_UPLOAD_ENABLE" == 1 ]]; then
    docker push "$TFCI_DOCKER_IMAGE"
  fi
fi

      const std::vector<EdgeSpec>& c_edges,  // expected ctrl edges
      bool is_exact_edges = true) {
    // Build a set of edges from fdef
    std::set<EdgeSpec> a_edges;  // actual edges
    // Get edges from inputs to body nodes and between body nodes
    for (const NodeDef& node_def : fdef.node_def()) {
      for (int i = 0; i < node_def.input_size(); ++i) {

  // While there is no immediate plan to deprecate dtype and shape in favor
  // of only using full type type information, this is a future possibility.
  //
  // Note that map_dtype_to_child_of_tensor() from core/framework/types.h
  // can be used to set a FullTypeDef based on dtype in a derived class if
  // appropriate.
  virtual tensorflow::FullTypeDef FullType() const = 0;

/// implements. Each of them is in a 1-to-1 correspondence with the wrapper
/// structures from the first section: these tables only contain function
/// pointers that operate on the corresponding data. Thus, the first argument of
/// each of these functions is a pointer to the paired struct and this argument
/// can be used to track state in between calls (from an object oriented point
/// of view, this can be viewed as a "vtable" for a "class" -- that is the

# calls javac with options that create generated files under a
# bazel-out directory. Next, it archives the generated source files
# into a srcjar directly under the root. There doesn't appear to be a
# simple way to parameterize this from bazel, hence this helper to
# "normalize" the srcjar layout.
#
# Arguments:
#   src_jar - path to the original srcjar
#   dest_jar - path to the destination
# Returns:
#   None
function cp_normalized_srcjar() {

+++ b/misc/sys/cdefs.h
@@ -423,4 +423,14 @@
 # endif
 #endif
 
+/* Undefine (also defined in libc-symbols.h).  */
+#undef __attribute_copy__
+#if __GNUC_PREREQ (9, 0)
+/* Copies attributes from the declaration or type referenced by
+   the argument.  */
+# define __attribute_copy__(arg) __attribute__ ((__copy__ (arg)))
+#else
+# define __attribute_copy__(arg)
+#endif
+
 #endif	 /* sys/cdefs.h */

            CACHEBUSTER=${{ steps.date.outputs.DATE }}
          tags: |
            gcr.io/tensorflow-sigs/build:${{ github.event.number }}-${{ matrix.python-version }}
          cache-from: |
            type=registry,ref=tensorflow/build:latest-${{ matrix.python-version }}
            type=registry,ref=gcr.io/tensorflow-sigs/build:${{ github.event.number }}-${{ matrix.python-version }}
          cache-to: type=inline

  virtual AbstractTensorInterface* CreateTensor(
      DataType dtype, absl::Span<const int64_t> dim_sizes) = 0;

  typedef void (*MemoryReleaser)(void* data, size_t len, void* arg);

  // Create a tensor instance from the given data buffer and description.
  // `memory_releaser` will be called on destruction, and it's responsible for
  // cleaning up the underlying buffer.
  virtual AbstractTensorInterface* CreateTensor(

  const size_t max_entries_;

  /// The callback to read timestamps.
  std::function<uint64_t()> timer_seconds_;

  /// Guards access to the cache and the LRU list.
  absl::Mutex mu_;

  /// The cache (a map from string key to Entry).
  std::map<std::string, Entry> cache_ ABSL_GUARDED_BY(mu_);

  /// The LRU list of entries. The front of the list identifies the most
  /// recently accessed entry.

# shared object directly.
#
# For example, from Python tf.load_op_library loads a custom op library (via
# dlopen() on Linux), the library finds libtensorflow_framework.so (no
# filesystem search takes place, since libtensorflow_framework.so has already
# been loaded by pywrap_tensorflow) and registers its ops and kernels via
# REGISTER_OP and REGISTER_KERNEL_BUILDER (which use symbols from