為Native Client平台提供轉譯的原生程式庫

Abstract

若要將一個程式庫跑在Native Client 沙盒平台上，則它的編碼必須符合一連串此沙盒平台的規範。一般來說，這些程式庫是透過，Native Client 所提供的編譯器工具來進行生成。這也是為何原生於原生平台上的原生程式庫無法直接在Native Client 上直接運行的主要原因。 若要將原生平台上的二元碼轉換成Native Client 可以使用的版本，則它必須將特定的某些指令轉換成符合Native Client 所規範的等價指令。此外，某些特定的指令必須透過nop (0x90)指令來推移到特定的位址上。然而上述的轉換操作可能會造成短定址的jump 指令位址空間不足使用，因此它們必須被轉換成大小較大的長定指模式。轉換成長定址的jump 指令又可能因為指令長度變長而導致其他的短定址jump 定址空間不足而需要轉換。直接地進行這些短指 令轉換可能會破壞之前轉換好的等價指令或導致某些特定指令偏離了指定的位址，甚至可能產生過長的運算時間。針對以上的問題我們提出一個演算法來解決。這個演算法會透過這些jump 指令的相依性來遞迴執行。 根據實驗結果，在網路頻寬小於7.5Mbps 時採用轉譯程式庫來取代下載程式庫來可減低程式的下載時間。轉譯過的程式庫中，檔案大小較小的程式庫約有平均8%左右的大小成長；檔案大小較大的程式庫則約有平均44%左右的大小成長。透過我們提出的演算法，約有平均約有8.4%左右的短定址的jump 指令在轉譯過程被轉換，針對較小型的程式庫，我們約有3~4%的效能損失；針對大小較大的程式庫，我們約有13~15%的效能損失。

rules. It replaces certain instructions with equivalent ones that follow the sandbox rules. In particular, some instructions must be placed at specific alignments and instructions cannot sit across the bundle boundaries. The above two constraints create "holes" in the instruction address space, which are padded with the nop (0x90) instructions. Furthermore, due to the padding process, some short jump instructions are space-exhausted and must be replaced with longer ones. However, naively replacing short jumps with longer ones might break the above two constraints and cause unacceptable time and space wasting. Therefore, we designed an algorithm for converting the spaceexhausted short jumps and guarantees that above two constraints hold. The algorithm works recursively through the dependency of the short jump instructions. By our experiment, small libraries will be larger than NaCl ones by about 8%. The normal sized libraries will be larger than NaCl ones by about 44% on average. We also suggest using the translator as an alternative to download the original libraries from the server if the network bandwidth is lower than 7.5Mbps. By algorithm, the change ratio of the short jumps is about 8.4%. There are 3~4% overhead on small libraries and 13~15% overhead on normal sized libraries.