In current superscalar processors, all floating-point resources are idle during the execution of integer programs. As previous works show, this problem can be alleviated if the floating-point cluster is extended to execute simple integer instructions. With minor hardware modifications to a conventional superscalar processor, the issue width can potentially be doubled without increasing the hardware complexity. In fact, the result is a clustered architecture with two heterogeneous clusters. We propose to extend this architecture with a dynamic steering logic that sends the instructions to either cluster. The performance of clustered architectures depends on the inter-cluster communication overhead and the workload balance. We present a scheme that uses run-time information to optimise the trade-off between these figures. The evaluation shows that this scheme can achieve an average speed-up of 35% over a conventional 8-way issue (4 int+4 fp) machine and that it outperforms the previously proposed one.

Peer Reviewed