arm64SpecialOperand = map[string]arm64.SpecialOperand{}
		for opd := arm64.SPOP_BEGIN; opd < arm64.SPOP_END; opd++ {
			arm64SpecialOperand[opd.String()] = opd
		}

		// Handle some special cases.
		specialMapping := map[string]arm64.SpecialOperand{
			// The internal representation of CS(CC) and HS(LO) are the same.
			"CS": arm64.SPOP_HS,
			"CC": arm64.SPOP_LO,
		}
		for s, opd := range specialMapping {
			arm64SpecialOperand[s] = opd

// tf_executor.island {_parallel_execution_ids=r0:0,p0:1} wraps "tf.OpD"
// tf_executor.island {_parallel_execution_ids=r0:1} wraps "tf.OpB"
// tf_executor.island {_parallel_execution_ids=r0:1,p0:0} wraps "tf.OpC"
// tf_executor.island {_parallel_execution_ids=r0:1,p0:1} wraps "tf.OpD"
//
// "tf.OpA" will not have `_parallel_execution_ids` attr,

        %c = "tf.opC"(%b) : (tensor<i32>) -> tensor<f32>
        tf_device.return %c, %b : tensor<f32>, tensor<i32>
      }) {device = "CPU:0"} : () -> (tensor<f32>, tensor<i32>)
      %d = "tf.opD"() : () -> tensor<i1>
      tf_executor.yield %a, %launch#0, %launch#1, %d : tensor<i1>, tensor<f32>, tensor<i32>, tensor<i1>
    }
    tf_executor.fetch
  }
  func.return
}

// CHECK-NOT: tf_executor.island

      tf_executor.yield %3 : tensor<i32>
    }
    %4:2 = tf_executor.island {
      %5 = "tf.opC"(%1#0, %arg1) : (tensor<i32>, tensor<i32>) -> tensor<i32>
      %6 = "tf.opD"(%5, %arg0) : (tensor<i32>, tensor<i32>) -> tensor<i32>
      tf_executor.yield %6 : tensor<i32>
    }
    tf_executor.fetch %4#0 : tensor<i32>
  }
  func.return %0 : tensor<i32>
}

// CHECK-SAME: ([[ARG_0:%.*]]: tensor<i32>)
// CHECK-NEXT: [[A:%.*]] = "tf.opA"([[ARG_0]])
// CHECK-NEXT: [[B:%.*]] = "tf.opB"([[A]])
// CHECK-NEXT: [[C:%.*]] = "tf.opC"([[A]])
// CHECK-NEXT: [[D:%.*]] = "tf.opD"([[C]])
// CHECK-NEXT: return [[B]], [[D]]
func.func @multiple_inner_ops(%arg0: tensor<i32>) -> (tensor<i32>, tensor<i32>) {
  %graph_result:2 = tf_executor.graph {
    %output0_0, %output0_1, %control0 = tf_executor.island {

        %2 = "tf.opA"() : () -> tensor<i1>
        %3 = "tf.opB"(%2) : (tensor<i1>) -> tensor<i1>
        tf_device.return %3 : tensor<i1>
      }, {
        %2 = "tf.opC"() : () -> tensor<i32>
        %3 = "tf.opD"(%2) : (tensor<i32>) -> tensor<i32>
        tf_device.return %3 : tensor<i32>
      }) : () -> (tensor<i1>, tensor<i32>)
      tf_executor.yield %1#0, %1#1 : tensor<i1>, tensor<i32>
    }

  "tf.TPUReplicateMetadata"() {_xla_compile_device_type = "TPU", _replication_info = "replicate", device = "/device:TPU:0", num_replicas = 1, topology = "topology"} : () -> ()
  %3 = "tf.opD"(%2) {_xla_compile_device_type = "TPU", _replication_info = "replicate", is_stateless = true} : (tensor<i1>) -> tensor<i1>
  %4 = "tf.opE"() {is_stateless = true} : () -> tensor<i1>

      %5 = "tf.opC"(%4) : (tensor<i1>) -> tensor<i1>
      tf_executor.yield %3, %5, %4 : tensor<i1>, tensor<i1>, tensor<i1>
    }
    %6:3 = tf_executor.island {
      %7 = "tf.opD"(%arg0) : (tensor<i1>) -> tensor<i1>
      %8 = "tf.opE"(%7) : (tensor<i1>) -> tensor<i1>
      tf_executor.yield %8, %7 : tensor<i1>, tensor<i1>
    }
    tf_executor.fetch %1#1, %1#0, %6#0 : tensor<i1>, tensor<i1>, tensor<i1>

func.func @replicate_with_multiple_operands() {
  %0 = "tf.opA"() : () -> tensor<*xi1>
  %1 = "tf.opB"() : () -> tensor<*xi1>
  %2 = "tf.opC"() : () -> tensor<*xi1>
  %3 = "tf.opD"() : () -> tensor<*xi32>
  %4 = "tf.opE"() : () -> tensor<*xi32>
  %5 = "tf.opF"() : () -> tensor<*xi32>
  %6 = "tf.opG"() : () -> tensor<*xf32>
  %7 = "tf.opH"() : () -> tensor<*xf32>
  %8 = "tf.opI"() : () -> tensor<*xf32>

        %c = "tf.opC"(%b) : (tensor<i32>) -> tensor<f32>
        tf_device.return %c, %b : tensor<f32>, tensor<i32>
      }) {device = "CPU:0"} : () -> (tensor<f32>, tensor<i32>)
      %d = "tf.opD"() : () -> tensor<i1>
      tf_executor.yield %a, %launch#0, %launch#1, %d : tensor<i1>, tensor<f32>, tensor<i32>, tensor<i1>
    }
    tf_executor.fetch
  }
  func.return
}

// CHECK:      %[[A:.*]] = "tf.opA"