Virtualizing register context.

Abstract

A processor designer may wish for an implementation to support multiple register contexts for several reasons: to support multithreading, to reduce context switch overhead, or to reduce procedure call/return overhead by using register windows. Conventional designs require that each active context be present in its entirety, increasing the size of the register file. Unfortunately, larger register files are inherently slower to access and may lead to a slower cycle time or additional cycles of register access latency, either of which reduces overall performance. We seek to bypass the trade-off between multiple context support and register file size by mapping registers to memory, thereby decoupling the logical register requirements of active contexts from the contents of the physical register file. Just as caches and virtual memory allow a processor to give the illusion of numerous multi-gigabyte address spaces with an average access time approaching that of several kilobytes of SRAM, we propose an architecture that gives the illusion of numerous active contexts with an average access time approaching that of a single active context using a conventionally sized register file. This dissertation introduces the virtual context architecture, a new architecture that virtualizes logical register contexts. Complete contexts, whether activation records or threads, are kept in memory and are no longer required to reside in their entirety in the physical register file. Instead, the physical register file is treated as a cache of the much larger memory-mapped logical register space. The implementation modifies the rename stage of the pipeline to trigger the movement of register values between the physical register file and the data cache. With the same size register file as a non register windowed machine, this architecture is within 1% of the execution time of an idealized register window machine. The virtual context architecture enables support for both register windows and simultaneous multithreading without increasing the size of the register file, increasing the performance by 50% over a single thread and 30% over a conventional multithreaded architecture.