Extensive Benchmarks Looking At AMD Znver1 GCC 9 Performance, EPYC Compiler Tuning

The LAME MP3 audio encoding performance does benefit from the -Ofast optimization level.

The PostgreSQL performance varied with -O2 with vectorization doing the best for read-only workloads while for a mix of reads/writes the best performance was exhibited with "-O3 -march=znver1".

For the last of these reference compiler benchmarks for a dual AMD EPYC server on the bleeding-edge GCC 9 compiler is the geometric mean of all the C/C++ benchmarks carried out that successfully ran all tested combinations. At the end of the day for the dozens of benchmarks that have been running the past week on this Dell PowerEdge R7425 server, the fastest performance was with "-O3 -march=znver1" going for the more aggressive "-Ofast" level did help some workloads but also regressed others while introducing the risk of unsafe math and lack of standards compliance. Likewise, LTO paid off well for some benchmarks but didn't help out some of the other benchmarks depending upon the structure of the code-base for the program under test. LTO also quickly drove up build times. For those having to deal with debugging binaries, at least the "-Og" level captured much of the compiler optimizations while not detracting from the ability to debug the binaries.

That's the latest state of the AMD EPYC/Znver1 compiler performance on GCC 9.The Znver1 compiler support is largely mature at this point in the GNU Compiler Collection code-base while the upcoming GCC 9 release also introduces the initial "znver2" target for next-generation AMD processors.