Java 堆疊運算之指令階層平行度開發

Abstract

Java虛擬平台是一個堆疊架構。由於絕大多數的Java指令的執行都受限於堆疊架構的頂端，因此使得平行執行的機會變的相當的稀少。在這篇論文中，我們針對高速處理的超純量Java處理器利用指令前送與指令重整緩衝區相互配合的機制提出一個可以平行存取的機制來打破傳統JVM架構中堆疊存取的循序限制。每個Java指令相互之前的關係會在執行的時候利用我們設計的指令相依性檢查法檢查出來，且會從指令重整緩衝區或是利用指令前送的方式抓取它們的所需的運算元，最後再將其結果寫回指令重整緩衝區中。另外，針對複雜的記憶體存取指令部分，我們亦利用JVM針對不同記憶體定址模式或是不同指令型態下的指令不會有記憶體位置相互重複的特性，設計一分散式指令窗戶架構，來簡化記憶體相依性的檢查工作。進而使得一個超純量Java處理器可以比傳統堆疊架構的Java處理器再一個時脈週期內執行較多的指令。經過我們模擬測試的結果顯示，我們設計的超純量處理器的執行效能可以比Sun PicoJavaII 好上3.1倍。

Java Virtual Machine architecture is a stack-based architecture. Because most operands of Java arithmetic instructions are restricted to the top of stack, there is little possibility for parallel execution of Java bytecodes in traditional stack architecture. In this thesis, we use reorder buffer and data forwarding working together with the stack, to break the performance bottleneck due to stack operations. At runtime, the dependences of Java bytecode sequence are checked by using our proposed dependency check mechanism and instead of accessing the top of stack, most arithmetic instructions can get source operands from data forwarding or the reorder buffer, and write results to the reorder buffer. For the memory reference instructions, we use the feature that there is no memory reference dependency among bytecodes if the instruction type or addressing modes are different, to design a destributed instruction window and this simplies the memory reference dependency detections. Then, the superscalar Java processor can execute more instructions in parallel in a clock cycle than stack based Java processors. Simulation result shows that the superscalar Java processor could achieve the performance speedup of average 3.1 versus the PicoJava II Java processor.