Enhanced bitwidth-aware register allocation

Abstract

Embedded processors depend on register files for performance, just like general-purpose processors in desktop and server systems. However, unlike general-purpose processors, the power consumption of register files poses a significant challenge for embedded processors, making it desirable for embedded processors to use as few registers as possible. Past research has indicated the potential for leveraging bitwidth analysis and bitwidth-aware register allocation to reduce register usage in embedded applications. This paper makes the following contributions in evaluating and enhancing bitwidth-aware register allocation for embedded applications. First, we compare the Tallam-Gupta bitwidth analysis with an idealized limit study, and show significant opportunities for enhancements. Second, we show how bitwidth-aware register allocation can be enhanced by enhanced bitwidth analysis for scalar and array variables, and also by enhanced coalescing of variables. Third, we use our prototype implementation of bitwidth-aware register allocation in gcc to compare the number of dynamic spill load/store instructions resulting from a) bitwidth-unaware allocation, b) bitwidth-aware allocation, c) enhanced bitwidth-aware allocation, and d) ideal profile-driven bitwidth-aware allocation. Our results show that our enhancements can reduce the number of dynamic spill load/store instructions to between 3% and 27% of the number obtained from the Tallam-Gupta algorithm. © Springer-Verlag Berlin Heidelberg 2006.