// See the License for the specific language governing permissions and
// limitations under the License.

[[incremental_build]]
= Incremental build

An important part of any build tool is the ability to avoid doing work that has already been done.
Consider the process of compilation.

///
/// All pointers defined in this structure point to memory allocated by the DSO
/// using an allocator provided by core TensorFlow when calling `TF_InitPlugin`.
///
/// IMPORTANT: To maintain binary compatibility, the layout of this structure
/// must not change! In the unlikely case that a new type of file needs to be
/// supported, add the new ops and metadata at the end of the structure.

  }
}

// Output `shard_type`, which is the type of each shard, given `full_type`. If
// the full shape is (num_cores_per_replica * a, b, c), then the shard shape is
// (a, b, c). `context_op` is used for error reporting, in case of errors.
LogicalResult GetShardShapedType(Operation* context_op,
                                 int num_cores_per_replica, Type full_type,
                                 Type& shard_type) {

  optional string name = 1;

  // singularName is the singular name of the resource.  This allows clients to handle plural and singular opaquely.
  // The singularName is more correct for reporting status on a single item and both singular and plural are allowed
  // from the kubectl CLI interface.
  optional string singularName = 6;

  // namespaced indicates if a resource is namespaced or not.

  optional string name = 1;

  // singularName is the singular name of the resource.  This allows clients to handle plural and singular opaquely.
  // The singularName is more correct for reporting status on a single item and both singular and plural are allowed
  // from the kubectl CLI interface.
  optional string singularName = 6;

  // namespaced indicates if a resource is namespaced or not.

	// basis point at (sweepHeapLiveBasis, pagesSweptBasis). At
	// any given time, the system is at (gcController.heapLive,
	// pagesSwept) in this space.
	//
	// It is important that the line pass through a point we
	// control rather than simply starting at a 0,0 origin
	// because that lets us adjust sweep pacing at any time while
	// accounting for current progress. If we could only adjust

// when you downcast, you should use this macro.  In debug mode, we
// use dynamic_cast<> to double-check the downcast is legal (we die
// if it's not).  In normal mode, we do the efficient static_cast<>
// instead.  Thus, it's important to test in debug mode to make sure
// the cast is legal!
//    This is the only place in the code we should use dynamic_cast<>.
// In particular, you SHOULDN'T be using dynamic_cast<> in order to

// lastChanged and lastChecked "times" so unchanged predecessors can be
// skipped on after-the-first iterations.  Doing this allows extra
// iterations by the caller to be almost free.
//
// It is important to know that the set representation used for
// startState, endState, and merges can share data for two sets where
// one is a small delta from the other.  Doing this does require a

	// These two values are still kept separate however because they are
	// distinct conceptually, and in previous iterations of the pacer the
	// distinction was more important.
	gcBackgroundUtilization = 0.25

	// gcCreditSlack is the amount of scan work credit that can
	// accumulate locally before updating gcController.heapScanWork and,

  // cost function, nor can we look at global information while deciding on
  // individual edges to contract.  Instead, we will make decisions on these
  // important edges then make decisions on all other edges, causing the highest
  // chance of all most important edges to be contracted.
  //
  // An example of where this might occur is with a digraph:
  // {A -> B, B -> C, A -> X, X -> C} where B is a Size operation and X is