Modulo Scheduling with Isomorphic Control Transformations

This dissertation addresses the complexities involved with scheduling in the presence of conditional branches. This is a particularly important problem for processors that execute multiple operations per cycle and are not fully utilized by local scheduling techniques. Since conditional branches introduce multiple execution paths, it is diicult for a global scheduler to keep track of the various paths and to select the appropriate operations to schedule. A new approach t o global instruction scheduling is presented that uses Isomorphic Control Transformations ICTs. If-conversion is used to convert an acyclic control ow graph into a large basic block o r h yper-block. Local scheduling techniques which are well-known and widely supported can then be applied to schedule operations. After scheduling, the control ow graph is regenerated using Reverse If-Conversion. One well-known local scheduling based technique is Modulo Scheduling. Modulo Scheduling is a software pipelining technique that eeectively schedules loops for high-performance processors. This dissertation highlights the beneets of Modulo Scheduling over other software pipelining techniques based on global scheduling. The ICTs are applied to Modulo Scheduling to schedule loops with conditional branches. Experimental results show that this approach allows more exible scheduling and thus better performance than Modulo Scheduling with Hierarchical Reduction. Modulo Scheduling with ICTs targets processors with no or limited support for conditional execution such as superscalar processors. However, in processors that do not require instruction set compatibility, support for Predicated Execution can be used. This dissertation shows that Modulo Scheduling with Predicated Execution has better performance iii and lower code expansion than Modulo Scheduling with ICTs on processors without special hardware support. iv DEDICATION To m y family, new and old. v ACKNOWLEDGMENTS First and foremost, I would like t o a c knowledge my advisor, Professor Wen-mei W. Hwu, for his intellectual, nancial, professional, and emotional support. I value his advice greatly and look forward to applying what I have learned from him as I start my career as a professor. Next, I would like t o a c knowledge the support of my Ph.D. committee members. Throughout my y ears in graduate school, Professor Michael Loui has provided me with valuable guidance. Furthermore, I have a deep respect for his commitment to excellence in education. I would also like t o a c knowledge Professors Janak Patel and David Kuck for helping me to focus my research. Finally, I w ould like t o a c …