Improving static branch prediction in a compiler

An ILP (Instruction-Level Parallelism) compiler uses aggressive optimizations to reduce a program's running time. These optimizations have been shown to be effective when profile information is available. Unfortunately, users are not always willing or able to profile their programs. A method of overcoming this issue is for an ILP compiler to statically infer the information normally obtained from profiling. This paper investigates one aspect of this inference: the static prediction of conditional-branch direction. The goals of this work are to utilize the source-level information available in a compiler when performing static branch prediction, to identify static-branch-prediction cases in which there is a high confidence that a branch will go in one direction at run time, to gain an intuitive understanding into the reasons why the static-branch-prediction heuristics are effective, and ultimately to improve the accuracy of the static branch prediction. The effectiveness of the static-branch-prediction heuristics developed in this paper is demonstrated on a set of programs from SPEC CINT92, SPEC CINT95, and the IMPACT compiler.