Branch Analysis And Optimization In Dynamic Binary Translation

Dynamic Binary Translation (DBT) is a technology that translates the instructions from the source instruction set architecture (ISA) to the target ISA at runtime. DBT is widely researched and used to support legacy binary code, provide program instrumentation, improve code performance, and so on. The performance of a DBT system is always an essential issue. And a major source of overhead is the execution of the branch instructions in source code, because it requires extra instructions to resolve the target address.This thesis first collects the data information of the branch instructions from the runtime environments of the benchmarks, and then it analyzes the categories of the branch instructions and their rates of a program at runtime. After that this thesis gives a simple description of all kinds of branch instructions handling schemes widely used in the DBT systems, and evaluates them base on the experiments performance on our own DBT system CrossBit. Then this thesis will introduce several branch instruction handling methods optimization applied in CrossBit, which is a multiple sources and multiple targets process level virtual machine. The optimization methods include stub handler scheme for the direct branch instructions, a combination of inline and Hash table method to the indirect branch instructions, and the improvements for the Return cache scheme to the return instructions.After the branch instruction handling scheme optimization, the performance of CrossBit is improved, and slow down to the native execution of the SPEC 2000 Int benchmarks is about 1.4~2.0 times, which is acceptable in DBT system crossing heterogeneous platforms.