Final report for Ceniit project 01.06 (2001–2007) Optimization Algorithms for Integrated Code Generation

In this project we developed integrative methods for generating high-quality code for embedded and instruction-level parallel processor architectures, including regular architectures such as superscalars and VLIW, as well as irregular architectures, such as clustered VLIW and DSP processors. Even though today’s DSP processors are more compiler friendly than a decade ago or than today’s network processors, they impose intricate structural constraints on the parallel use of functional units, the placement of operands in registers, and the scheduling of instructions. Additionally, these various constraints strongly depend on each other. As a consequence, traditional approaches to code generation that solve the problem by treating the subproblems in isolation in separate, subsequent phases, can either not be applied at all or produce code of very poor quality. On the other hand, DSP code must meet very high quality demands in terms of execution time, required memory space, code size, and energy consumption. Often, the application code is fixed, its time-critical part is comparably small in size, and aggressive overnight or parallel optimization is feasible. However, an integrated approach to optimal code generation that considers instruction selection, instruction scheduling and register allocation simultaneously constitutes a very hard combinatorial optimization problem. We developed and implemented exact algorithms, based on dynamic programming and on integer linear programming, that can solve the integrated problem for reasonable problem sizes and realistic architectures. For larger problem instances we provided heuristic algorithms. Beyond execution time, we also targeted alternative code optimization goals such as energy consumption. The algorithms are implemented in our retargetable code generation framework called OPTIMIST, where the target processor architecture is modeled in our architecture description language xADML.