Compiler and microarchitecture mechanisms for exploiting registers to improve memory performance.

摘要

The performance of the memory hierarchy has become one of the most critical elements in the performance of desktop, workstation and server computer systems. This is in large part due to the growing gap between memory and processor speed and the large number of memory operations present in typical programs. This problem is called the memory bottleneck. There are a number of techniques for containing this problem, including register allocation, caching, and prefetching. These techniques, which either eliminate memory operations or reduce their apparent effect, can be implemented in either hardware or software. This dissertation presents combined hardware and software methods to implement and manage the register file more effectively, with the primary focus to eliminate memory operations.; This dissertation first examines the register requirements of integer codes and shows that a combination of advanced compiler optimizations combined with a large register file can improve performance up to 45%. The improvements show the importance of a large register file to integer codes. Although a large number of registers is good from the compiler's perspective, a large register file can be slow and expensive to build. This dissertation proposes a new caching technique that retains the advantages of a large register file without the speed and cost penalties.; Studies of existing compiler optimizations showed that a common optimization called register promotion, which increases the utilization of the register file and reduces memory operations, is limited by static alias analysis in the compiler. This dissertation introduces a new compiler optimization called speculative register promotion and a new hardware structure called the store load address table to address this problem.; This dissertation also proposes several new metrics to measure compiler performance in units of hardware complexity and examines memory accesses in the context of their programming language source code in order to determine the relative importance and characteristics of these memory operations.; The dissertation concludes with an examination of a new research compiler for the C programming language called MIRV that was used to perform this research. Various aspects of the compiler and engineering trade-offs made during its design are explained.