Analysis Of LLVM Infrastructure And Backend Porting For ARCA3

Due to the mechanism of compiler porting, it is not necessary for micro-processor manufacturer to develop new compiler for every new-built hardware platform. And reusing the open-source compiler is a better choice, which means great significance for reserving the advantage of portable compiler, reducing the development costs and shortening the time to market.Based on the open-source LLVM compiler infrastructure developed by Illinois University and ARCA3 micro-processor with independent intellectual property rights of China, our issue ports the LLVM compiler to the ARCA3 micro-processor and consummates the compiler tool chains for ARCA3 design and development platform, using the advanced retargetable property of LLVM.The entire porting process can be divided into three stages: LLVM compiler infrastructure exploration, ARCA3 architecture exploration, LLVM back-end porting and system testing. First, we analyze the LLVM compiler infrastructure, and introduce different components of LLVM. Then we extract the designing methods of LLVM which is better than conventional compiler architecture, especially the backend porting mechanism. Secondly, we introduce the ARCA3 architecture and extract useful information for interfacing with LLVM, according to the data organization property, register group setting property, instruction set property, stack frame layout property and calling convention property. Finally, combined with LLVM backend porting mechanism and ARCA3 architecture, we realize the porting method and process, and then complete the system testing based on the LLVM test files on the assembly code level.Ultimately ported LLVM compiler supports most critical instructions of ARCA3, and also different types of functions. Upon completion of the transplant, not only the tool chains for ARCA3 design and development platform are consummated, the applications of ARCA3 are expanded, and the advanced features of LLVM can benefit ARCA3 platform too, improving its performance at compiler time as much as possible.