Check out the new USENIX Web site.

Online Call Graph Construction

Context-insensitive call graphs are commonly used by IPAs, where a method is represented as one node in the call graph. There exists a directed edge from a method A to a method B if A calls B.

Dynamic class hierarchy information can be used to build a conservative call graph at runtime. However, it is desirable to have a more precise call graph for most interprocedural analyses. We propose a new mechanism for profiling and constructing context-insensitive call graphs at runtime. The mechanism initializes call edges using a profiling code stub. When the code stub gets executed, it generates a new call edge event, then it triggers method compilation if the method is not compiled yet, and finally patches the address of the real target. The mechanism captures the first execution event of each call edge, and the first execution has some profiling overhead. The repeated calls only need to execute at most one more instruction. Clients, such as call graph builders, can register callback routines called by a profiling code stub when new call edges are discovered. Callbacks can perform necessary actions before the callee is invoked.

The remainder of this section is structured as follows. First, in Section 2.1, we briefly introduce a conservative approach for building call graphs using runtime class hierarchy information. In Section 2.2 we give the necessary background, describing the existing implementation of virtual method tables in Jikes RVM. In Section 2.3 we describe the basic mechanism we propose for building call graphs at runtime, and in Section 2.4 we show how this basic mechanism can be optimized to reduce overheads.

Feng Qian 2004-02-16