Efficient register and memory assignment for non-orthogonal architectures via graph coloring and MST algorithms

摘要

Finding an optimal assignment of program variables into registers and memory is prohibitively difficult in code generation for application specific instruction-set processors (ASIPs). This is mainly because, in order to meet stringent speed and power requirements for embedded applications, ASIPs commonly employ non-orthogonal architectures which are typically characterized by irregular data paths, heterogeneous registers and multiple memory banks. As a result, existing techniques mainly developed for relatively regular, orthogonal general-purpose processors (GPPs) are obsolete for these recently emerging ASIP architectures. In this paper, we attempt to tackle this issue by exploiting conventional graph coloring and maximum spanning tree (MST) algorithms with special constraints added to handle the non-orthogonality of ASIP architectures. The results in our study indicate that our algorithm finds a fairly good assignment of variables into heterogeneous registersand multi-memories while it runs extremely faster than previous work that employed exceedingly expensive algorithms to address this issue.