Colloquium Speaker

Speaker: Eduard Mehofer
University of Vienna
Topic:Feedback-Directed Optimization based on Probabilistic Data Flow Information
Date:Thursday, September 6, 2001
Time:11:00 AM
Place:Gould-Simpson, Room 701


Refreshments will be served in the 7th-floor lobby of Gould-Simpson at 10:45 AM


ABSTRACT


In feedback-directed optimization (FDO) runtime information is used to improve the efficiency of programs for typical program runs. Our approach to FDO is based on probabilistic data flow analysis (PDFA). PDFA systems determine the probabilities with which data flow facts will hold at some program point. Subsequent optimizing transformations take this probabilistic data flow information into account to achieve better results.

In this talk we give an overview of probabilistic data flow systems and present a novel probabilistic data flow framework which calculates data flow probabilities more accurately. Experiments based on GNU gcc and SPEC95 show that the approach is feasible and delivers significantly better results.