// CHECK: %[[BROADCAST_TO:.*]] = "tfl.broadcast_to"(%arg0, %[[BROADCAST_ARGS]]) : (tensor<8x7x6x5x?x3x2x1xf32>, tensor<8xi64>) -> tensor<8x7x6x5x?x3x2x1xf32>
  // CHECK: %[[BROADCAST_TO_1:.*]] = "tfl.broadcast_to"(%arg1, %[[BROADCAST_ARGS]]) : (tensor<?x3x2x1xf32>, tensor<8xi64>) -> tensor<8x7x6x5x?x3x2x1xf32>

// [1 x 1 x 1 x m*n] by the outermost reshape_op.
// problem: The inefficiency here is that the innermost reshape_op and the
// broadcast_op are introducing unnecessary leading and trailing 1s'.
// fix: Remove the unnecessary 1s' in the inner reshape_op and broadcast_op.
struct SqueezeReshapesAroundBroadcastOp
    : public OpRewritePattern<TFL::BroadcastToOp> {
  using OpRewritePattern<TFL::BroadcastToOp>::OpRewritePattern;

	broadcaster := events.NewBroadcaster(&events.EventSinkImpl{Interface: client.EventsV1()})
	ctx, cancel := context.WithCancel(context.Background())
	defer cancel()

	informerFactory := informers.NewSharedInformerFactory(client, 0)
	sched, err := New(
		ctx,
		client,
		informerFactory,
		nil,
		profile.NewRecorderFactory(broadcaster),
		WithProfiles(

// CHECK: %[[BROADCAST:.+]] = "tfl.broadcast_to"(%[[ARG0]], %[[SHAPE]]) : (tensor<1x2x!quant.uniform<i8:f32, 2.000000e+00:3>>, tensor<2xi32>) -> tensor<3x2x!quant.uniform<i8:f32, 2.000000e+00:3>>
// CHECK: return %[[BROADCAST]]

// -----

// Tests that a quantized `stablehlo.broadcast_in_dim` is converted to
// `tfl.transpose` and `tfl.broadcast_to` when `broadcast_dimensions` is not in
// ascending order.

  %4 = "tfl.broadcast_args"(%2, %3) : (tensor<8xi64>, tensor<8xi64>) -> tensor<8xi64>
  %5 = "tfl.broadcast_to"(%arg0, %4) : (tensor<8x7x6x5x?x3x2x1xi1>, tensor<8xi64>) -> tensor<8x7x6x5x?x3x2x1xi1>
  %6 = "tfl.broadcast_to"(%arg1, %4) : (tensor<8x7x6x5x?x3x2x1xf32>, tensor<8xi64>) -> tensor<8x7x6x5x?x3x2x1xf32>
  %7 = "tfl.broadcast_to"(%arg2, %4) : (tensor<?x3x2x1xf32>, tensor<8xi64>) -> tensor<8x7x6x5x?x3x2x1xf32>

    TFL::QConstOp rhs_qconst_op;

    auto GetBroadcastedConstOp = [&](Value operand) -> TFL::QConstOp {
      if (auto broadcast_op = dyn_cast_or_null<stablehlo::BroadcastInDimOp>(
              operand.getDefiningOp())) {
        auto stablehlo_const_op = dyn_cast_or_null<stablehlo::ConstantOp>(
            broadcast_op.getOperand().getDefiningOp());
        auto const_uniform_quantized_type =

by a scalar, broadcasting (usually) confers some time or space
benefit, as the broadcasted tensor is never materialized.

However, `broadcast_to` does not carry with it any such benefits.
The newly-created tensor takes the full memory of the broadcasted
shape. (In a graph context, `broadcast_to` might be fused to
subsequent operation and then be optimized away, however.)
  }];

  let arguments = (ins

      !llvm::dyn_cast_or_null<TFL::BroadcastToOp>(defining_op)) {
    return nullptr;
  }

  Value broadcast_shape = defining_op->getOperand(
      1);  // Broadcasted shape operand of BroadcastTo op.
  Operation* parent_of_defining_op = broadcast_shape.getDefiningOp();
  if (!llvm::dyn_cast_or_null<TF::BroadcastArgsOp>(parent_of_defining_op) &&
      !llvm::dyn_cast_or_null<TFL::BroadcastArgsOp>(parent_of_defining_op)) {

  %emptyf = "tf.Cast"(%emptyi) : (tensor<i32>) -> tensor<f32>
  %size_shape = "tf.Const"() {value = dense<[10, 8, 4]> : tensor<3xi32>} : () -> tensor<3xi32>
  %tl = "tf.BroadcastTo"(%emptyf, %size_shape) : (tensor<f32>, tensor<3xi32>) -> tensor<10x8x4xf32>
  // TensorListPushBack lowering
  %index_in_list = "tf.Const"() {value = dense<0> : tensor<1xi32>} : () -> tensor<1xi32>

  func.return %r0, %r1 : tensor<0xi32>, tensor<3xi32>
}

// -----

// Test valid tf.BroadcastTo
// CHECK-LABEL: func @testBroadcastTo(%arg0: tensor<16xf32>)
func.func @testBroadcastTo(%arg0: tensor<16xf32>) -> tensor<16x16x16x16xf32> {
  %cst = arith.constant dense<16> : tensor<4xi32>
  %0 = "tf.BroadcastTo"(%arg0, %cst) : (tensor<16xf32>, tensor<4xi32>) -> tensor<16x16x16x16xf32>
  func.return %0 : tensor<16x16x16x16xf32>