} : (tensor<?x100352xf32>, tensor<100352x10xf32>) -> tensor<?x10xf32>
    %3 = "tf.Identity"(%2) {device = ""} : (tensor<?x10xf32>) -> tensor<?x10xf32>
    return %3 : tensor<?x10xf32>
  }

  // CHECK-LABEL: func.func private @composite_dot_general_fn_1
  func.func private @composite_dot_general_fn_1(%arg0: tensor<?x100352xf32>, %arg1: tensor<100352x10xf32>) -> tensor<?x10xf32> {
    // CHECK-NOT: tf.CustomAggregator

// CHECK-NOT: call

// -----

func.func @handles_mhlo_regions() {
  "glue.someop"() : () -> ()
  %4 = mhlo.constant dense<0.0> : tensor<200x10xf32>
  %13 = mhlo.constant dense<-0.0> : tensor<f32>
  %29 = mhlo.reduce(%4 init: %13) applies mhlo.add across dimensions = [1] : (tensor<200x10xf32>, tensor<f32>) -> tensor<200xf32>
  return
}

// CHECK: func @handles_mhlo_regions
// CHECK: call [[name:@[^(]*]](

// -----

module {
  func.func @same_scale_ptq_test(%arg0: tensor<*xf32>) -> tensor<*xf32> {
    %cst = arith.constant dense<[-1, 144]> : tensor<2xi32>
    %cst_1 = arith.constant dense<1.0> : tensor<144x10xf32>
    %cst_2 = arith.constant dense<0.1> : tensor<10xf32>
    %0 = "quantfork.stats"(%arg0) {
      layerStats = dense<[-1.28, 1.28]> : tensor<2xf32>
    } : (tensor<*xf32>) -> tensor<*xf32>
    %1 = "tf.MaxPool"(%0) {

// CHECK: return %[[identity]] : tensor<1x3x2x2xf32>

func.func @cast_bf16_matmul_to_fp32(%arg0: tensor<1x10xf32>) -> (tensor<1x2xf32>) {
  %cst = "tf.Const"() {device = "", value = dense<1.000000e+01> : tensor<10x2xbf16>} : () -> tensor<10x2xbf16>
  %0 = "tf.Cast"(%arg0) {Truncate = false, device = ""} : (tensor<1x10xf32>) -> tensor<1x10xbf16>

    // CHECK-NEXT: return %[[DIV]] : tensor<4x10xf32>
    %0 = "tf.Softsign"(%arg0) : (tensor<4x10xf32>) -> tensor<4x10xf32>
    func.return %0 : tensor<4x10xf32>
}

// -----

// CHECK-LABEL: func @softsign_grad
func.func @softsign_grad(%arg0: tensor<4x10xf32>, %arg1: tensor<4x10xf32>) -> tensor<4x10xf32> {

    // CHECK-NEXT: %[[ONE:.*]] = mhlo.constant dense<1.000000e+00> : tensor<f32>

  // expected-error @+1 {{'tfl.prelu' op result type '10x10' not broadcast compatible with broadcasted operands's shapes '10x10x10x10'}}
  %0 = "tfl.prelu"(%arg0, %arg1) : (tensor<10x10x10x10xf32>, tensor<10x10x10x10xf32>) -> tensor<10x10xf32>
  func.return %0 : tensor<10x10xf32>
}

// -----

  %11 = "mhlo.dot"(%10, %5) : (tensor<32x3920xf32>, tensor<3920x10xf32>) -> tensor<32x10xf32>
  %12 = "mhlo.broadcast_in_dim"(%2) <{broadcast_dimensions = dense<1> : tensor<1xi64>}> : (tensor<10xf32>) -> tensor<32x10xf32>
  %13 = mhlo.add %11, %12 : tensor<32x10xf32>
  %14 = mhlo.maximum %13, %0 : tensor<32x10xf32>
  return %14 : tensor<32x10xf32>
}
// CHECK: ReadVariableOp
// CHECK: mhlo.convolution
// CHECK: mhlo.add

    func.return %1 : tensor<2xf32>
  }

  // CHECK-LABEL: binary_op_broadcast
  func.func @binary_op_broadcast(%arg0: tensor<4x1xf32>, %arg1: tensor<4x1x4xf32>) -> tensor<4x4x4xf32> {
    // CHECK: %[[BROADCAST0:.*]] = "mhlo.broadcast_in_dim"(%arg0) <{broadcast_dimensions = dense<[1, 2]> : tensor<2xi64>}> : (tensor<4x1xf32>) -> tensor<4x4x1xf32>
    // CHECK: %[[RESHAPE0:.*]] = mhlo.reshape %[[BROADCAST0]] : (tensor<4x4x1xf32>) -> tensor<4x4xf32>

  %max_size = "tf.Const"() {value = dense<10> : tensor<i32>} : () -> tensor<i32>
  // CHECK-NOT: tf.EmptyTensorList
  // CHECK: tf.BroadcastTo
  // CHECK-SAME: tensor<10x16xf32>
  %tl0 = "tf.EmptyTensorList"(%element_shape, %max_size) : (tensor<1xi32>, tensor<i32>) -> tensor<!tf_type.variant<tensor<16xf32>>>
  // CHECK-NOT: tf.TensorListReserve
  // CHECK: tf.BroadcastTo

    }
  }
  return new_arg_type;
}

// Combination of value producer and port of value produced (e.g.,
//   <value result output>:<value in output tensor>,
// so for tf.Const -> tensor<10x20xf32>, [0,2,18] would point to a unique output
// scalar value).
struct ValuePort {
  PointerUnion<Operation*, BlockArgument> producer;
  SmallVector<unsigned int, 2> port;