具折疊功能之Java Bytecode的指令層次平行度分析

Abstract

堆疊機器的效率主要被堆疊循序存取限制住。昇陽公司 (Sun Microelectronics) 在設計picoJava堆疊處理器時，利用了一項機制--堆 疊指令折疊--來增進效率。除了昇陽公司的折疊法則，還有另外兩種折疊 的技巧，分別是: 固定樣式的折疊與POC組態。於本論文中，我們將研究 影響指令平行行度的因素以及在使用指令折疊技巧--POC組態--後所能萃 取出的平行度。我們提出了五個影響指令平行度的因素，分別是: 堆疊指 標更名、局部變數更名、有限運算、有限排程視窗以及有限折疊個數。由 統計數據顯示，與POC組態比較我們可獲得兩倍的指令平行度；若沒有堆 疊指標更名，則我們將只能獲得一半的理論上限的平行度。更藉由實驗， 我們可發現每個平行指令群大約須四個運算就可達理論上限值。對指令視 窗而言: 相對於無限指令視窗大小，32位元組的指令視窗大小可達82%指 令平行度；而64位元組的指令視窗大小可達93%指令平行度。最後我們可 得到一個結論:當未來的設計朝向高階的機器時，折疊運算與堆疊指標更 名是不可或缺的機制。 The performance of a stack machine is limited mainly by sequential stackacceses. A performance enhancement mechanism - Stack Operations Folding -was used in Sun Microelectronics' picoJava design. There are another twofolding techniques: fixed- pattern folding and POC model beyond Sun'.In thisthesis, we use the Java bytecode language as the target machine language, analyze the factors affecting ILP, and extract parallelism among bytecodesafter folding using POC model.Five factors: stack pointer renaming, localvariable renaming, finite operations, finite scheduling window size andlimited foldable numbers were simulated and evaluated. Statistical data showthat the ILP can be doubled as compared to original figure of POC model. Andwithout stack pointer renaming, the performance gain is only half oftheoretical upper bound. From the experiments, we observe that 4 operationsfor each PIG (Parallel Instruction Group) is enough to approximate thetheoretical upper bound. With a 32-byte or a 64-byte window size, theperformance can reach 82% or 93% as compared to the infinite window size.Finally, we conclude that stack operations folding and stack pointer renamingare the most important mechanisms in designing higher-end stack machines.