Identifying Multiply-Add Operations in Kylin Compiler

摘要

There has been a trend in the embedded microprocessor industry in which the Instruction Set Architecture of a processor includes multiply-add (MADD) instruction to offer potentially smaller code size and/or better performance for specific classes of numerical applications. To exploit the power of this instruction, compilers should be able to identify MADD opportunities and use general optimization techniques to generate code sequence with MADD instructions. This paper reports our MADD identify strategies implemented in the Kylin C compiler. We developed two MADD opportunity identification algorithms which are implemented in different phases of the Kylin compiler. One of this methods is implemented in the middle end based on a MADD super tree. The other method is implemented in the back end based on CGIR representation. Our strategies can process up to four MADD-candidate patterns. Therefore, it can identify almost all of the MADD opportunities. We also discuss the pros and cons of implementing the MADD identify strategy in two different phases of a compiler in terms of code size and performance.