Global Address Register Allocation for Array References in DSPs [abstract] (PDF, PostScript)
Guilherme de Lima Ottoni
Master's Thesis, Institute of Computing, Universidade Estadual de Campinas (UNICAMP), December 2002.
Best Master's Thesis Award (2nd Prize Nationwide).

The technological advances in computing systems have stimulated the growth of the embedded systems market, which is continuously becoming more ordinary in people's lives, for example in mobile phones, palmtops and automotive control systems. Because of their characteristics, these new applications demand the combination of low cost, high performance and low power consumption. One way to meet these constraints is through the design of specialized processors. However, processor specialization imposes new challenges to the development of software for these systems. In particular, compilers -- generally responsible for code optimization -- need to be adapted in order to produce efficient code for these new processors.

In the digital signal processing arena, such as in cellular telephones, specialized processors, known as DSPs (Digital Signal Processors), are largely used. DSPs typically have few general purpose registers and very restricted addressing modes. In addition, many DSP applications include large data streams processing, which are usually stored in arrays. As a result, studing array reference optimization techniques became an important task in compiling for DSPs. This work studies this problem, known as Global Array Reference Allocation (GARA). The central GARA subproblem consists of determining, for a given set of array references to be allocated to the same address register, the minimum cost of the instructions required to keep this register with the correct address at all program points. In this work, this subproblem is modeled as a graph theoretical problem and proved to be NP-hard. In addition, an efficient algorithm, based on dynamic programming, is proposed to optimally solve this subproblem under some restrictions. Based on this algorithm, two techniques to solve GARA are proposed. Experimental results, from the implementation of these techniques in the GCC compiler, compare them with previous work in the literature. The results show the effectiveness of the techniques proposed in this work.