Aligned Scheduling: Cache-Efficient Instruction Scheduling for VLIW Processors

The performance of statically scheduled VLIW processors is highly sensitive to the instruction scheduling performed by the compiler. In this work we identify a major deficiency in existing instruction scheduling for VLIW processors. Unlike most dynamically scheduled processors, a VLIW processor with no load-use hardware interlocks will completely stall upon a cache-miss of any of the operations that are scheduled to run in parallel. Other operations in the same or subsequent instruction words must stall. However, if coupled with non-blocking caches, the VLIW processor is capable of simultaneously resolving multiple loads from the same word. Existing instruction scheduling algorithms do not optimize for this VLIW-specific problem. We propose Aligned Scheduling, a novel instruction scheduling algorithm that improves performance of VLIW processors with non-blocking caches by enabling them to better cope with unpredictable cache-memory latencies. Aligned Scheduling exploits the VLIW-specific cache-miss semantics to efficiently align cache misses on the same scheduling cycle, increasing the probability that they get serviced simultaneously. Our evaluation shows that Aligned Scheduling improves the performance of VLIW processors across a range of benchmarks from the Mediabench II and SPEC CINT2000 benchmark suites up to 20%.