Compiler Optimization Techniques for Scheduling and Reducing Overhead

Exploiting parallelism in loops in programs is an important factor in realizing the potential performance of processors today. This dissertation develops and evaluates several compiler optimizations aimed at improving the performance of loops on processors. An important feature of a class of scientific computing problems is the regularity exhibited by their access patterns. Chapter 2 presents an approach of optimizing the address generation of these problems that results in the following: (i) elimination of redundant arithmetic computation by recognizing and exploiting the presence of common sub-expressions across different iterations in stencil codes; and (ii) conversion of as many array references to scalar accesses as possible, which leads to reduced execution time, decrease in address arithmetic overhead, access to data in registers as opposed to caches, etc. With the advent of VLIW processors, the exploitation of fine-grain instruction-level parallelism has become a major challenge to optimizing compilers. Fine-grain scheduling of inner loops has received a lot of attention, little work has been done in the area of applying it to nested loops. Chapter 3 presents an approach to fine-grain scheduling of nested loops by formulating the problem of finding theminimum iteration initiation interval as one of finding a rational affine schedule for each statement in the body of a perfectly nested loop which is then solved using linear programming. Frequent synchronization on multiprocessors is expensive due to its high cost. Chapter 4 presents a method for eliminating redundant synchronization for nested loops. In nested loops, a dependence may be redundant in only a portion of the iteration space. A characterization of the non-uniformity of the redundancy of a dependence is developed in terms of the relation between the dependences and the shape and size of the iteration space. Exploiting locality is critical for achieving high level of performance on a parallel machine. Chapter 5 presents an approach using the concept of affinity regions to find transformations such that a suitable iteration-to-processor mapping can be found for a sequence of loop nests accessing shared arrays. This not only improves the data locality but significantly reduces communication overhead