Global Trade-oo between Code Size and Performance for Loop Unrolling on Vliw Architectures

Many media processors 28, 7, 14, 8, 18, 27], used for computing intensive embedded applications, are VLIW architectures that rely on the compiler to exploit Instruction Level Parallelism. Loop unrolling is generally used to expose instruction parallelism but computing the unrolling factor is very diicult as instruction cache misses and spill code can cancel the expected beneet of the transformation. Moreover increasing the code size directly impacts on the embedded system cost. In this paper, we propose a method, called UFC (Unrolling Factor computation under Constraints) to compute unrolling factors of set of loops while taking into account code size, a major issue for embedded systems.