Interactive Source-Level Debugging of Optimized Code (PostScript version, PDF version)
Le-Chun Wu
Phd thesis, Department of Computer Science, University of Illinois, Urbana, IL, August 1999

With an increasing number of executable binaries generated by optimizing compilers today to fully utilize advanced architecture features, it has become a necessity to support debugging optimized code. One of the most difficult problems in debugging globally optimized code is to recover the expected variable values at source breakpoints. To solve this problem, the debugger not only has to stop the execution at appropriate places to preserve necessary program state, but also needs to be able to correctly associate storage locations with source variables.

In this dissertation, a new framework for debugging globally optimized code is proposed. This framework consists of a novel breakpoint implementation scheme and a new data location tracking mechanism. In the proposed breakpoint implementation scheme, the debugger takes over the control of execution early and executes instructions under a new forward recovery model. This enables the debugger to recover the expected behavior of a program even in the presence of optimization. Also the source breakpoints are reported to the user in the order specified by the original source program and the behavior of exceptions meets what the user expects.

The new data location tracking scheme keeps track of variable definition information during optimization. A data-flow analysis based on the definition information preserved is performed to collect data that is then used to generate the run-time data location information. With this data location information, a debugger incorporating the proposed breakpoint implementation scheme can determine if the expected value of a variable is available at a source breakpoint and how to recover it.

The debugging framework has been prototyped in both the IMPACT compiler and an experimental debugger. Experiments conducted on several integer benchmark programs have yielded encouraging results. The overhead in executable file size and compile time incurred by this framework is reasonable. Compared with previous work, the proposed approach is much more effective in the recovery of the expected variable values.

[ IMPACT Main Page | Team Members | Publications | Software | FAQ ]