Accurate static estimators for program optimization

  • Authors:

  • Tim A. Wagner;Vance Maverick;Susan L. Graham;Michael A. Harrison

  • Affiliations:

  • Computer Science Division, University of California, Berkeley;Computer Science Division, University of California, Berkeley;Computer Science Division, University of California, Berkeley;Computer Science Division, University of California, Berkeley

  • Venue:
  • PLDI '94 Proceedings of the ACM SIGPLAN 1994 conference on Programming language design and implementation
  • Year:
  • 1994

Quantified Score

Hi-index 0.00

Visualization

Abstract

Determining the relative execution frequency of program regions is essential for many important optimization techniques, including register allocation, function inlining, and instruction scheduling. Estimates derived from profiling with sample inputs are generally regarded as the most accurate source of this information; static (compile-time) estimates are considered to be distinctly inferior. If static estimates were shown to be competitive, however, their convenience would outweigh minor gains from profiling, and they would provide a sound basis for optimization when profiling is impossible.We use quantitative metrics to compare estimates from static analysis to those derived from profiles. For C programs, simple techniques for predicting branches and loop counts suffice to estimate intraprocedural frequency patterns with high accuracy. To determine inter-procedural estimates successfully, we combine function-level information with a Markov model of control flow over the call graph to produce arc and basic block frequency estimates for the entire program.For a suite of 14 programs, including the C programs from the SPEC92 benchmark suite, we demonstrate that static estimates are competitive with those derived from profiles. Using simple heuristics, we can determine the most frequently executed blocks in each function with 81% accuracy. With the Markov model, we identify 80% of the frequently called functions. Combining the two techniques, we identify 76% of the most frequently executed call sites.