Extracting fine-grain profiles of in-order executions of instruction level parallel programs

Optimizing compilers targeted to instruction level parallel (ILP) architectures schedule program instructions in such a way so as to minimize the number of execution stalls, called bubbles, that occur during program execution because of hazards. These bubbles are estimated by compilers on the basis of the target processor functional model. Unfortunately, these functional models are often inaccurate, especially in retargetable compilers, and thus lead the compiler to imprecise estimations. This paper presents a technique that “detects” the bubbles by observing program execution in a set of preliminary runs, instead of estimating them. The applications of this technique range from performance evaluation of programs and efficiency analysis of functional models to novel techniques of profile-driven code optimization. The paper presents the technique and shows its applicability in a case study and a set of experiments