Building Process Virtual Machine Via Dynamic Binary Translation

Traditional computer software is closely dependent on the underlying machine platform, which means program compiled for one machine cannot run on another. It makes software uneasy to migrate and mobilize, and makes hardware designer unable to change the hardware/software interface for further innovation because of the importance of backward compatibility. However, virtual machine breaks the binding of software and hardware, serving as an intermediate layer in between. Binary Translation, which translates the source machine code to target, is the key techology that implements virtual machine.The work of this thesis is to study and build virtual machine via dynamic binary translation technology. After summarizing binary translation and related technologies, we describe an infrastructure of dynamic binary translation, named CrossBit, which translates application from SimpleScalar to x86.The design goals of CrossBit are retargetability and extensibility. The essential of retargetability is the intermediate language - VInst, a low-level virtual machine instruction set. VInst effectively separates source and target machine, and minimizes the machine-dependent part of the system. CrossBit also values modularity and sound interface definition, which makes it easy to extend and suitable for the purpose of research prototyping. This thesis discusses the system architecture of CrossBit, especially the rationale of VInst design, and analyzes individual modules. Finally, experiment shows that CrossBit runs more efficiently than traditional software interpreter.