Research And Implementation Of The Autovectorization For YHFT-Matrix Complier

In recent years, single instruction multiple data (SIMD) extensions, which offer higherperformance, lower energy for multimedia and Digital signal processing, have been incorporatedinto many microprocessors. Among the ways to make use of SIMD extensions, the best way forthe pragammers is autovectorization, which can reduce much burden on pragammers, reduce thecosts of application development, and improve the security and portability of the applications.YHFT-Matrix is an architecture with SIMD features. The number of SIMD instructionsaccounts for almost50%in the YHFT-Matrix instruction set. Therefore, the autovectorization ofSIMD instructions is a vital part of YHFT-Matrix complier. We designed and implemented theautovectorization of YHFT-Matrix SIMD instructions with the open source complier GCC.The work of this paper include:The autovectorization approach for YHFT-Matrix was designed and implemented on thebase of the current research of the autovectorization as well as the mechanism of vectorization inGCC. The YHFT-Matrix back-end was supported in GCC, and the mapping of simple SIMDoperations was realized.Further more, a lot of optimization work was done to deal with the specifc SIMDinstructions of YHFT-Matrix as follows:First, a scheme of data type extension was proposed for the saturation arithmetic ofYHFT-Matrix. In this way, the saturation arithmetic’s intermediate representation (IR) can besimplified.Second, the special multiplication operation in our architecture was supported by modifyingUsual Arithmetic Conversion of C standard and by this way, the superfluous statement in IR wasremoved, and the input to the vectorier was simplified.Third, an approach based on operation extension was presented for the special mathematicaloperations which were mapped by means of library functions in GCC making these operations inthe loop can be translated into relevant SIMD instructions of YHFT-Matrix directly.Finally, since the vectorizer was lack of capabilities in vectorization, a technology usingpragma dierectives to control the vectorier was raised and implemented. The programmers canuse the dierectives to pass information to the vectorier in vectorization process.We implemented and verified the work described above on GCC-4.3.2platform, and theresults show that the correct SIMD instructons are obtained in the mapping of simple operationswhile the capabilities of the vectorizer is improved for YHFT-Matrix by using optimizationtechnologies in this paper, and the autovectorization approach for YHFT-Matrix raised by thispaper is proved to be correct and effective.