}

    def "broadcasts settings evaluated events to closures"() {
        given:
        def called = false
        def closure = { called = true }

        when:
        gradle.settingsEvaluated(closure)

        and:
        gradle.buildListenerBroadcaster.settingsEvaluated(null)

        then:
        called
    }

    def "broadcasts projects loaded events to closures"() {
        given:

     */
    public void addAll(Collection<? extends T> listeners) {
        broadcast = broadcast.addAll(listeners);
    }

    /**
     * Adds a {@link Dispatch} to receive events from this broadcast.
     */
    public void add(Dispatch<MethodInvocation> dispatch) {
        broadcast = broadcast.add(dispatch);
    }

    /**
     * Adds an action to be executed when the given method is called.

// Creates a pass that hoists reads out of a replicate that are on a variable
// whose value is broacast to all replicas.
std::unique_ptr<mlir::OperationPass<mlir::func::FuncOp>>
CreateHoistBroadcastReadPass();

// Creates a pass that moves broadcasts from TF host ops to XLA code, encoded as
// XlaAllReduces. This enables use of the device network for broadcasts, which
// is faster.
std::unique_ptr<mlir::OperationPass<mlir::func::FuncOp>>

  void RegisterExecution(const NameAttrList& function);

  // Registers a cluster compilation. Increments the compilation count and
  // accumulates the compile time for the given cluster. Also broadcasts an
  // XlaJitCompilationActivity.
  virtual Status RegisterCompilation(const NameAttrList& function,
                                     int64_t compile_time_us,

#include "xla/mlir_hlo/mhlo/IR/hlo_ops.h"

namespace mlir {
namespace odml {

// Broadcasts the 1D value tensor 'value_1d' to the shape of 'result_type'. If
// 'shape_value' is initialized, creates a dynamic broadcast, otherwise creates
// a static broadcast.
Value broadcastToFeatureDim(Location loc, RankedTensorType result_type,
                            Value value1d, Value shape_value,

//
// Examples:
//
//	"RU_SAE.Z" => {rounding: 3, zeroing: true}
//	"Z" => {zeroing: true}
//	"BCST" => {broadcast: true}
//	"SAE.Z" => {sae: true, zeroing: true}
type evexSuffix struct {
	rounding  byte
	sae       bool
	zeroing   bool
	broadcast bool
}

// Rounding control values.
// Match exact value for EVEX.L'L field (with exception of rcUnset).
const (

  Value all_reduce =
      CreateAllReduce(replicate, inner_builder, launch.getResult(0));

  orig_to_new[val_bcast] = all_reduce;
  return success();
}

// Move all suitable broadcasts across replicas to the `cluster` into the
// `cluster`.
LogicalResult MoveAllBroadcastsToCluster(ClusterOp cluster,
                                         ReplicateOp replicate,

        // Non-parallel tensors from _EagerConst/tf.constant are implicitly
        // broadcast, i.e. set as the input to each parallel operation. This
        // allows code like "tf.constant(1.)" or "tf.reduce_sum(..., axis=1)"
        // (where the value starts on the host), without allowing other implicit
        // copies/broadcasts. Other implicit copies may be supported eventually,

}

def XlaBroadcastPass : Pass<"tf-xla-broadcast", "mlir::func::FuncOp"> {
  let summary = "Moves a broadcast from host into XLA, encoded as XlaAllReduce";

  let description = [{
    This pass moves brodcasts from host TF ops into XLA. This enables use of
    the inter-device network, which is faster than the inter-host network.
    Broadcasts in XLA are encoded as XlaAllReduce. An all_reduce with all 0

func.func @broadcast_shift_right_unsigned(%arg0: tensor<4xui8>, %arg1: tensor<2x4xui8>) -> tensor<2x4xui8> {
  // CHECK: %[[BROADCAST:.*]] = "mhlo.broadcast_in_dim"(%arg0) <{broadcast_dimensions = dense<1> : tensor<1xi64>}> : (tensor<4xui8>) -> tensor<2x4xui8>
  // CHECK:  mhlo.shift_right_logical %[[BROADCAST]], %arg1 : tensor<2x4xui8>
  %0 = "tf.RightShift"(%arg0, %arg1) : (tensor<4xui8>, tensor<2x4xui8>) -> tensor<2x4xui8>