Instruction scheduling using bitvectors for VLIW architecture

P. Shankar
Abstract:We extend a framework for retragetable code generation with the following additions. Firstly, the front-end for the framework which involves writing appropriate specification for parsing Lcode generated by the IMPACT front-end[1] has been extended to accept the latest version of Lcode, and validated on some large benchmarks. Secondly,we have implemented the wavefront scheduling technique [2] which uses bitvectors for path based data dependency representation proposed recently for scheduling for VLIW architectures. We have integrated the wavefront scheduler with the Lcode translator and code generator generator tool (CGG) [3]. The Lcode translator generates expression Directed Acyclic Graphs(DAGs), and the code generator which is output by CGG from the machine specifications in terms of a regular tree grammar, generates the data dependency DAGs from a traversal of the expression DAGs for each of the basic blocks. The wavefront scheduler schedules these data dependency DAGs. Current work described below will target the TMS320C62xx VLIW architecture [4].
What problem does this paper attempt to address?