Abstract
In this thesis, we describe two techniques to optimize hardware stack machine performances on Java environments. Our hardware model basically is a Java processor with a fixed-size stack cache, which directly implements the method frame activation allocations of a software Java virtual machine (JVM). In the first category, we present a technique to solve the problem about the stack allocations for intra-procedural methods in the Java processor. We put emphasis on the issue with the local variable allocation of the method frame. A structure or object unfolding technique can be used to transform heap accesses into stack accesses. The composite object in Java is accessed via heap in memory, while object unfolding transforms heap access to scalar and stack references. For Java processors with a fixed-size stack cache, unlimited employment of structure unfolding techniques however will result in the size of local variables excelling the size of the stack cache. Thus it will reduce the performance gains. For example, ITRI-made Java processor experiences performance penalty in this scenario. To solve this problem, we propose a mechanism, reverse object-unfolding, to report an allocation scheme for a given size of the stack allocation according to our cost model. In the second category of this problem, we also extend our framework for stack allocations to deal with inter-procedural cases. We model this problem into equations and propose a heuristic algorithm based on the domain decomposition of the call graphs of a program to solve the stack allocation problems for inter-procedural cases. Our solution deals with the reduction of both memory references and stack flushes.Our experiment is performed on the ITRI-made Java processor architecture and Kaffe VM simulator. The ITRI-made Java processor is with a fixed-size stack cache and directly allocates the method frames on the stack cache. We report experimental results and profiling data for both intra-procedural and inter-procedural cases. The experiments indicate our proposed methods are promising in speedup Java programs on the Java processor with a fixed-size stack cache.