Check out the new USENIX Web site.


Introduction

Interprocedural analyses (IPAs) derive more precise program information than intraprocedural ones. Static IPAs provide a conservative approximation of runtime information to clients for optimizations. A foundation of IPA is the call graph of the analyzed program. IPAs for Object-Oriented (OO) programs share some common challenges. Virtual calls (polymorphism) make call graph construction difficult. Further, since the code base tends to be large, the complexity and precision of the analysis must be carefully balanced.

One difficulty of call graph construction for OO languages lies in how to approximate the targets of polymorphic calls. In addition to polymorphism, call graph construction for Java is further complicated by the presence of dynamic class loading. Static IPAs assume that the whole program is available at analysis time. However, this may not be the case for Java. A Java program can download a class file from the network or other unknown resources. Even when all programs exist on local disks, a VM may choose to load classes lazily, on demand, to reduce resource usage and improve responsiveness [21]. When a JIT compiler encounters an unresolved symbolic reference, it may choose to delay resolution until the instruction is executed at runtime. A dynamic analysis has to deal with these unresolved references. A more subtle problem, usually ignored by static IPAs for Java, is that a runtime type is defined by both the class name and its initial class loader. Therefore, a correct dynamic IPA has to be incremental (dealing with dynamic class loading), efficient, and type safe.

Although Java's dynamic features pose difficulties for program analyses, there are many opportunities at runtime that can only be enjoyed by dynamic analyses. For example, a dynamic analysis only needs to analyze loaded classes and invoked methods. Therefore, the analyzed code base can be much smaller than in a conservative static analysis. Further, dynamic class loading can improve the precision of type analyses. The set of runtime types can be limited to loaded classes. Thus, a dynamic analysis has more precise type information than its static counterpart. Further, in contrast to the conservative (pessimistic) nature of static analysis, a dynamic one can be optimistic about future execution, if used in conjunction with runtime invalidation mechanisms [23,18,12,30].

Over the last 10 years, VM technology has greatly advanced. JIT compilers now implement most intraprocedural data-flow analyses that can be found in static compilers [1,22]. Further performance improvements have been achieved using adaptive and feedback-directed compilation [4,5].

Dynamic interprocedural analysis, however, has not yet been widely adopted. Some type-based IPAs [18,23] have gained ground in JIT compilation environments. However, work relating to more complicated, reachability-based IPAs, such as dynamic points-to analysis and escape analysis, is only just starting to emerge [15].

In this paper, we present a call graph construction mechanism for reachability-based interprocedural analyses at runtime. Instead of approximating a call graph as in static IPAs, our mechanism uses a profiling code stub to capture invoked call edges. The mechanism overcomes difficulties caused by dynamic class loading and polymorphism. Most overhead happens at JIT compilation and class loading time. It has only small overhead on the performance of applications in a JIT environment. A very desirable feature of the mechanism is that call graphs can be built incrementally while execution proceeds. This enables speculative optimizations using runtime invalidations for safety.

Dynamic IPAs seem more suitable for long-running applications in adaptive recompilation systems. Pechtchanski and Sarkar [23] described a general approach of using dynamic IPAs. A virtual machine gathers information about compiled methods and loaded classes in the initial state, and performs recompilation and optimizations only on selected hot methods. When the application reaches a ``stable state'', information changes should be rare.

Based on our new runtime call graph mechanism, we describe the design and implementation of an online version of an example IPA, XTA type analysis [32]. Dynamic XTA uses dependency databases to handle unresolved types and field references. The analysis is driven by VM events such as compilation, class loading, or the discovery of new call edges.

The rest of paper is organized as follows. Section 2 introduces our new call graph construction mechanism which serves as the basis for dynamic IPAs. In the following section, Section 3, we describe the design of a specific dynamic IPA, dynamic XTA type analysis, in the presence of lazy class loading. The call graph mechanism and dynamic XTA have been implemented in Jikes RVM. Section 4 analyzes the cost of call graph profiling and compares the characteristics of profiled call graphs to conservative ones built by dynamic CHA on a set of standard benchmarks. Section 5 discusses related work and conclusions are presented in Section 6.

Feng Qian 2004-02-16