Architectural Support for Java Constant Pool Resolution

Abstract

在Java程式執行當中，解釋constant pool中用符號表示之事物的動作是天生循序、且需花費很多的週期數和記憶體存取才能完成。因此，若將有效地執行解釋動作，將可大幅提升Java的程式執行效能。在這篇論文中，我們提出的縮減解釋時間之方法是將先前解釋引例物件相關指令所獲得的有用資訊存到硬體緩衝器上，利用參考實物資料表(referenced entity info table)與代理表(handle table)以縮減解釋時間，並有效地處理物件更新位置之問題。透過Trace-driven的方式對SPECjvm98和BenchBean標杆程式進行方案模擬。結果顯示我們所提出的方案在與昇陽所提出的改寫指令方案相較下，可加速解釋1.17倍，且不需修改指令空間。

The operations of resolution of symbolic reference in the Java constant pool are inherently serial and require many clock cycles and memory references, and hence efficient resolution is critical to the execution speed of Java code. Our approach of improving the resolution performance is to cache the useful information of previous resolutions of instance object related bytecode in hardware buffers. In this thesis, referenced entity info table (REIT) and handle table (HT) that can reduce the resolution time while maintaining the efficiency of object relocation, are proposed. A detailed trace-driven simulation of the proposed method on SPECjvm98 benchmark and BenchBean test program shows that our proposed method accelerates resolution by 1.17x while enable modify-less of instruction space, with respect to the rewriting method that proposed by Sun.