Check out the new USENIX Web site.


Related Work

Static call graph construction for OO programming languages focuses on approximating a set of types that a receiver of a polymorphic call site may have at runtime. Static class hierarchy analysis (CHA) [9] treats all subclasses of a receiver's declaring class as possible types at runtime. Rapid type analysis (RTA) [6] prunes the type set of CHA by eliminating types that do not have an allocation site in the whole program. Static CHA and RTA examine the entire set of classes. Propagation-based algorithms propagate types from allocation sites to receivers of polymorphic call sites along a program's control flow. Assignments, method calls, field and array accesses may pass types from one variable to another. Context-insensitive algorithms can be modelled as unification-based [29] or subset-based [3] propagation as points-to analysis. The complexity varies from $O(N\alpha(N,N))$ for unification-based analysis to $O(N^3)$ for subset-based analysis. Context-sensitive algorithms [10] might yield more precise results but are difficult to scale to large programs. Since CHA and RTA do not use control flow information, both are considered to be fast algorithms when compared with propagation-based algorithms. Both VTA [31] and XTA analysis [32] are simple propagation-based type analyses for Java. The analyses can either use a call graph built by CHA/RTA, then refine it, or build the call graph on the fly [25].

Ishizaki et. al. [18] published a new method of utilizing class hierarchy analysis for devirtualization at runtime. If the target of a virtual call is not overridden in the current class hierarchy, a compiler may choose to inline the target directly with a backup code of normal virtual call. To cope with dynamic class loading, the runtime system monitors class loading events. If a newly loaded class overrides a method that has been directly inlined in some callers, the code of callers has to be patched with the backup path before class loading proceeds. Pechtchanski and Sarkar [23] presented a framework for performing dynamic optimistic interprocedural analysis in a Java virtual machine. Similar to dynamic CHA, their framework builds detailed dependencies between optimistic assumptions for optimizations and runtime events such as method compilation. Invalidation is a necessary technique for correctness when the assumption is invalidated. Neither of these approaches explored reachability-based analysis which requires a call graph as the base. Our work inherits the merits of their work, supporting optimistic optimizations and invalidations. Bogda and Singh [8] experimented an online interprocedural shape analysis, which uses an inlining cache to construct the call graph at runtime. However, their implementation was based on bytecode instrumentation, which incurs a large overhead. Our work aims to build an accurate call graph with little overhead to enable reachability-based IPAs at runtime.

In parallel to our work, Hirzel et. al. [15] adapted a static subset-based pointer analysis to a runtime analysis in Jikes RVM. In their work, an approach similar to ours is used to handle lazy class loading and unresolved method references. However, they used a conservative call graph constructed by dynamic CHA, and the analysis also considers the dataflow in a method. It would be interesting to see how much the smaller call graph produced by our mechanism could improve their pointer analysis results.

Code-patching [18] and stack-rewriting[12,17] are necessary invalidation techniques for optimistic optimizations. Those operations might be expensive at runtime. An optimization client should use these techniques wisely. For example, if an optimistic optimization has rare invalidations, these techniques can be applied. In situations of frequent invalidations or incomplete IPA information, an optimization may choose runtime checks to guard optimized code.

Static IPAs for Java programs assume whole classes are available at analysis time. Dynamically loaded classes should be supplied to the analysis manually. Sreedhar et.al. [28] proposed an extant analysis framework which performs unconditional static optimizations on references that can only have types in the closed world (known classes by analysis), and guided optimizations on references with possible dynamically loaded types. However, the effectiveness of online extant analysis may be compromised by the laziness of class loading at runtime. Java poses access restrictions on fields by modifiers. Field analysis [13] uses access modifiers to derive useful properties of fields for optimizations.

A new wave of VM technology is adaptive feedback-directed optimizations [16,22,4]. Sampling is a technique for collecting runtime information with low costs. Profiling information provides advice to compilers to allocate resources for optimizing important code areas. Compared with feedback-directed optimizations, optimizations based on dynamic IPAs can be optimistic using invalidation techniques instead of using runtime checks. Dynamic IPAs also provide a complete picture of an executing program. The new proposed mechanism is capable of finding all invoked call edges in executed code. In many cases, profiling information can be aggregated with IPAs. For example, Jikes RVM's adaptive system samples call stacks periodically and builds a weighted, partial call graph for adaptive inlining. A complete call graph constructed by our mechanism could be annotated with sampled weights on edges and clients could perform probabilistic analysis using the call graph.

Feng Qian 2004-02-16