Analyzing the individual/combined effects of speculative and guarded execution on a superscalar architecture

Speculative execution is a technique by which instructions are executed before the condition that controls it is evaluated. This can increase the performance if some of the idle CPU cycles are now used to execute speculated instructions. Guarded execution is a technique in which the branch instruction is eliminated and control dependencies are converted to data dependencies. This can help reduce some of the side-effects involved with branch instructions besides creating larger compilation units. However, excessive application of either one of them can result in dismal performance. Conventional approaches have used a one-time feedback metric and made all decisions based on it. We present a new way of designing feedback metrics and show how it can be used to regulate the effects of dynamic speculation and the side-effects of applying guarded execution statically. The proposed method presents 0.3 to 0.6-fold improvements over a conventional scheme using SPEC benchmarks