LLVM Clang 6.0 Benchmarks On AMD's EPYC Yield Some Performance Benefits
SQLite is slightly faster when built under Clang 6.0...
While in the Polybench benchmark the performance had slowed down under the newer compiler.
But optimized FFTW meanwhile is showing a nice 7% improvement in performance with LLVM Clang 6.0.
While the TSCP chess benchmark and other common workloads were seeing about the same performance or less than a 2% difference.
John The Ripper is faintly slower on Clang 6.0.
The Himeno pressure solver workload is almost 10% faster with LLVM Clang 6.0.
The Bullet Physics Engine was trying to pull a bit ahead with Clang 6.0 but didn't make much of a difference in the grand scheme.
Overall across the span of many benchmarks run, only in a few of the cases was there a measurable change in performance either way. FFTW and Himeno were among the big winners with znver1 on Clang 6.0 where they were 7~10% faster than the current stable Clang 5.0 compiler release.
Additional results can be found via this OpenBenchmarking.org result file.