AMD AOCC 1.2 Code Compiler Offers Some Performance Benefits For EPYC

On this EPYC system, the LLVM-based compilers were the fastest with the FLAC audio encoding test.

But when it comes to LAME MP3 encoding, GCC remained slightly ahead.

AOCC 1.2 did have some narrow first place finishes with the Redis NoSQL server.

While for the NGINX web server test, all four tested compilers yielded roughly the same performance.

When AOCC 1.0 was released last year shortly after the Ryzen/EPYC processor launch, there wasn't much to it and we had a hard time finding performance changes compared to the upstream LLVM/Clang state at the time. Fortunately, AOCC 1.1 and 1.2 have managed to pull in some more optimizations to better differentiate it as an LLVM/Clang downstream. AOCC 1.2 in several different benchmarks were leading over Clang 6.0 by up to a few percent. Another benefit now with AOCC 1.2 is that it makes FLANG easily available as the LLVM-based Fortran compiler. I'll be trying out FLANG on EPYC in a follow-up article. It's great that AMD continues investing in Linux compiler optimizations and hopefully more of these AOCC tuning/optimization changes will wind up in upstream LLVM in the not too distant future.