Dynamic Dependency Collapsing

In this dissertation, we explore the concept of dynamic dependency collapsing. Performance increases in computer architecture are always introduced by exploiting additional parallelism when the clock speed is fixed. We show that further improvements are possible even when the available parallelism in programs are exhausted. This performance improvement is possible due to executing instructions in parallel that would ordinarily have been serialized. We call this concept dependency collapsing. We explore existing techniques that exploit parallelism and show which of them fall under the umbrella of dependency collapsing. We then introduce two dependency collapsing techniques of our own. The first technique collapses data dependencies by executing two normally dependent instructions together by fusing them. We show that exploiting the additional parallelism generated by collapsing these dependencies results in a performance increase. Our second technique collapses resource dependencies to execute instructions that would normally have been serialized due to resource constraints in the processor. We show that it is possible to take advantage of larger in-processor structures while avoiding the power and area penalty this often implies