High-Level Optimizations for Newly Created and Existing Microprocessor Architectures

Abstract

This work is devoted to the description of a compiler designed based on LLVM. The LLVM infrastructure has an integrated UTL library containing implementation of analysis and optimization algorithms -- auto-parallelizing and vectorization in particular. Using benchmarks from the SPEC/CPU2006 test suite, the obtained solution is compared with a Linaro GCC compiler on an ARMv8 architecture multi-core processor and with GCC and Clang on a dual-core MIPS processor. The transformations implemented in the UTL allowed achieving a significant performance gain. Besides, the work describes experience of adding back-end to LLVM for a new DSP architecture Lynx developed by the Milandr RC.