Research On Modeling And Optimization Technique Of Binary Translation

With the continuous development of electronic information technology,the independent research and popularization of domestic processor architecture platforms have become an important factor to guarantee national security and to improve social production efficiency.The efficient transplantation method can provide good software support for the new processor platform to improve the breadth and depth of the platform applying.But there is some inheritance software without source code which hinders the transplantation of this kind of software on the new platform and greatly limits the application and popularization of the new architecture.As a cross-platform software transportation technology,binary translation technology can realize the cross-structure execution of binary programs between different architecture with software methods.When the architecture of source platform and target platform is very different,the difference has a great influence on the correctness and efficiency of binary translation.The complexity of the application also has a significant effect on the translation.Improving the function and the performance of the translator is the main objective of the binary translation studies.In the binary translation transplantation research on the SW processor platform which is applied on the Sunway Taihu Light computing system.(The computing system won the first place of TOP500 in the last three times.)Through the comparison and analysis of lots of practice on binary translation,this dissertation constructs an abstract model of binary translation process.This model abstracts the core functions of the translation system and can effectively describe the similarities and differences between dynamic and static translation pattern.Then it provides theoretical guidance for further research on the improvement of translation efficiency.The main works and creations are as the following:(1)A unified model of binary translation process had been built.Based on the analysis of similarity and otherness,the main functions in processes of dynamic and static binary translation can be abstracted as a transformation of the instruction sequence and executing order maintaining two parts.Extending the formalized model of machine simulation,this dissertation built the model which describes the binary translation process as the transformation from one instruction sequence set to another.It was presented as the corresponding basis to expand the function and improve the performance of binary translation technique.(2)A framework of dynamic binary translation with static pre-translation was presented with analysis of the dynamic binary translator cost in runtime.The process of dynamic binary translation can be divided into query,translation and execution three stages.After analyzing the executing process in runtime of dynamic binary translator,the method of pre-translating part of the instruction sequence out of runtime was devised.The static pre-translation in this translation framework does not take up the dynamic runtime overhead.Save translating time of the corresponding instruction sequences in runtime can improve the whole performance of the translator.(3)The query process in dynamic binary translator with lots static information was optimized.Both using static pre-translation method and using local library function replacement will affect the query overhead,especially when the static information is too much.Combined with the characteristics of these methods,the information used in the dynamic query process can be optimized in static.The problem of declining efficiency in dynamic query process had been addressed with the optimized method and data structure.(4)A register allocation optimization method based on the priority of storage unit mapping is proposed.The existing binary translator generally uses the intermediate representation to implement local code generation.In code generation,the corresponding register allocation is required for the variables in the intermediate representation.However,variables in the intermediate representation cannot be distinguished in the general register allocation scheme.This dissertation distinguished the intermediate variable with the register using frequency of the source platform.The register allocation strategy guided by the statistics reduces the redundancy of load and store instruction and improves the efficiency of the generating code.(5)A binary translation framework oriented to MPI parallel program was proposed.With the popularization of parallel programming,binary translation technology also had to face with the problems of various parallel applications.However,the MPI parallel application is essentially based on an API provided by a set of MPI libraries to implement various parallel interactions.In this dissertation,a binary translation framework for MPI parallel application was devised.The framework can make the translator provide source applications MPI-API service using local MPI library function and make the transportation has good efficiency.Model,methods and related algorithms proposed in this dissertation have been implemented and applied to the translator developed by our research group.The correctness and validity of the model and algorithms have been verified through the experiment.