Resarch On Cross-Plateform Compiler Analysis And Optimization Technique Based On Peak Architecture

In recent years,due to the development of the semiconductor industry at home and abroad,the domestic development in the field of chips has been booming.Because of the boom in hardware technology,the variety of traditional chips and artificial intelligence chips has been diversified,and traditional compilers have evolved towards custom compilers.The compiler is the bridge between the computer platform and the application,and the compiler correctly and efficiently compiles and converts the application source code into executable programs based on the machine platform's parameter information.During the compilation process,different optimization options are often overlaid to optimize the application in different ways.The ability of the compiler to optimize is influenced by the machine platform parameters and the architecture characteristics of its own compiler.In compiler optimization work,the architecture combination formed by the compiler and the machine platform is used as a base object for analysis.Due to the difficulty of compiler optimization work,the choice of reference architecture combinations has often been weakened in previous work.But only in the case of multi-architecture combinations can the target compiler be optimized to provide the maximum possible performance boost.But increasing the number of architecture combinations tends to make testing tasks heavy,data analysis heavy,and optimization difficult.To this end,this paper proposes a cross-platform compiler analysis technique based on peak architecture,which reduces the workload of compiler performance analysis and the difficulty of compiler optimization.The main research work of this thesis is summarized in the following sections.(1)A cross-platform compiler analysis and optimization technique based on peak architecture is proposed.The technique uses four steps,namely two-dimensional dynamic testing,peak architecture analysis,coarse-grained advantage optimization positioning,and fine-grained core differentiation analysis,to analyze the gap between the target compiler and the reference compiler in detail and propose a detailed optimization plan.(2)Using the SPEC CPU2017 as a test case and gcc,icc,aocc,and llvm as analytical compilers,and gcc as a target compiler on the Intel platform and llvm as a target compiler on the AMD platform,the practicality and accuracy of this technology was verified.(3)The source code logic structure and the main functional approach of the gcc compiler are analyzed from three aspects: front-end,mid-end(optimizer)and back-end of the compiler.Using this technique as a guide,the gcc compiler on the Intel platform was provided with several directions to be optimized,including quantization optimization,prefetch optimization,and loop unfold optimization,which resulted in a 67% performance improvement in the prefetch optimization direction.