CreateTPUClusterFormationPass(bool strict_clusters = false);

// Creates a pass that extracts outside compilation (Host ops inside device
// cluster) at head/tail of Device cluster to run before/after XLA computation.
std::unique_ptr<mlir::OperationPass<mlir::ModuleOp>>
CreateExtractHeadTailOutsideCompilationPass();

// Creates a pass that extract outside compilation (Host ops inside cevice

// it's double-declared (it would cost extra in the common case to verify
// this). But the MethodSet computation will not find it due to the name
// collision caused by the double-declaration, leading to an internal
// inconsistency while we are verifying one computation against the other.
// var _ = T1c{}.Pointer

// T2's method declared before the type.
func (*T2) f /* ERROR "field and method" */ () {}

{{DECLS_FROM_OBJ_FILE}}

{{NS_START}}
// {{CLASS}} represents a computation previously specified in a
// TensorFlow graph, now compiled into executable code. This extends the generic
// XlaCompiledCpuFunction class with statically type-safe arg and result
// methods. Usage example:
//
//   {{CLASS}} computation;
//   // ...set args using computation.argN methods
//   CHECK(computation.Run());

// use_tuple_args: when this is true, always create a tuple argument for the
//   entry computation.
// enable_op_fallback: when this is true, prefer tf2xla fallback kernels over
// MLIR
//   native kernels for legalization to HLO.
// return_tuple: when this is true, always create a tuple result for the
//   entry computation.
// shape_determination_fns: Contains layout preference fn and shape

// Creates a pass that lifts operations on external resource variables from
// device computation nested in `tf_device::LaunchOp` out so that resource
// variable load operations are all before device computation while resource
// variable store operations are all after device computation. After this pass,
// device computation no longer interacts with external resource variables.

    }
  }
}

// Since we have the outputs from host and device computation after moving
// outside compiled ops, we can create the actual parallel_execute regions.
// Still, one region is for the host computation for outside compilation and
// the other one is for the original Device cluster computation.
mlir::tf_device::ParallelExecuteOp CreateFinalParallelExecuteOp(

      %computation = "tf_device.cluster_func"(%read) {func = @computation, use_spmd_for_xla_partitioning = true} : (tensor<i32>) -> tensor<i32>
      "tf.AssignVariableOp"(%partitioned_variable, %computation) : (tensor<!tf_type.resource<tensor<i32>>>, tensor<i32>) -> ()
      return
    }

    func @computation(%arg0: tensor<i32>) -> tensor<i32> {
      return %arg0: tensor<i32>

    $mlir_module is a serialized MLIR module with a `main` function that contains target computation.
    $metadata is a serialized TPUCompileMetadataProto describing the shapes and types of the inputs to the computation, as well as a mapping onto the TPU pod topology.

      location, output_type, concat_dimension_op.getOutput(), inputs);
}

// For tile sharded inputs to TPU computation, inject split op between the
// input values and TPU computation so that tiled input values are passed in
// as inputs to TPU computations. If more than one dimension is sharded, then
// a tree of connected split ops are added before tf_device.parallel_execute op.

		go func() {
			once.Do(onceBody)
			done <- true
		}()
	}
	for i := 0; i < 10; i++ {
		<-done
	}
	// Output:
	// Only once
}

// This example uses OnceValue to perform an "expensive" computation just once,
// even when used concurrently.
func ExampleOnceValue() {
	once := sync.OnceValue(func() int {
		sum := 0
		for i := 0; i < 1000; i++ {
			sum += i
		}
		fmt.Println("Computed once:", sum)