Improving Software Pipelining by Hiding Memory Latency with Combined Loads and Prefetches

Modern processors and compilers hide long memory latencies through non-blocking loads or explicit software prefetching instructions. Unfortunately, each mechanism has potential drawbacks. Non-blocking loads can significantly increase register pressure by extending the lifetimes of loads. Software prefetching increases the number of memory instructions in the loop body. For a loop whose execution time is bound by the number of loads/stores that can be issued per cycle, software prefetching exacerbates this problem and increases the number of idle computational cycles in loops. In this paper, we show how compiler and architecture support for combining a load and a prefetch into one instruction, called a prefetching load, can give lower register pressure like software prefetching and lower load/storeunit requirements like non-blocking loads. On a set of 106 Fortran loops we show that prefetching loads obtain a speedup of 1.07–1.53 over using just non-blocking loads and a speedup of 1.04–1.08 over using software prefetching. In addition, prefetching loads reduced floating-point register pressure by as much as a factor of 0.4 and integer register pressure by as much as a factor of 0.8 over non-blocking loads. Integer register pressure was also reduced by a factor of 0.97 over software prefetching, while floating-point register pressure was increased by a factor of 1.02 versus software prefetching in the worst case.