iWatcher: Simple and General Architectural Support for Software Debugging

Recent impressive advances in microprocessor performance have failed to deliver significant gains in ease of software debugging. This is a major shortcoming of the state of the art, given that software bugs have major implications on computer system reliability and programmer time. Specifically, software bugs account for as much as 40% of computer system failures [10] and cost the U.S. economy $59.5 billion annually, or 0.6% of the GDP [11]! Code debugging is currently largely done using software techniques. One approach is to perform checks statically [2, 4, 5, 18]. Most static tools require significant involvement of the programmer to write specifications or annotate programs. In addition, most static tools are limited by aliasing problems and other compile-time limitations, especially for C/C++ programs. As a result, many bugs often remain in the software even after aggressive static checking. A second approach is to monitor execution dynamically. Examples of dynamic monitors include Purify [7], Valgrind [15], the Intel thread checker [9], DIDUCE [6], Eraser [14], CCured [3, 12], and other tools [1]. The strength of this approach is that the analysis is based on actual execution paths and accurate values of variables and aliasing information. However, most dynamic checkers suffer from two limitations. First, they are often computationally expensive. Some dynamic checkers slow down a program by 6-30 times [6, 14], which makes such tools undesirable for production runs. Moreover, some timing-sensitive bugs may never occur with these slowdowns. Second, most dynamic checkers rely on compilers or pre-processing tools to insert instrumentation and, therefore, are limited by imperfect variable disambiguation. Consequently, some accesses to a monitored location may be missed by the instrumentation tool. Because of this reason, some bugs are caught much later than when they actually occur, which makes it hard to find the root cause of the bug. The following C code gives a simple example.