LLVM Clang 3.4 SVN Compiler Optimization Level Tests

For the multi-threaded C-Ray ray-tracing benchmark there were significant performance improvements between -O0 and -O1 and then going from -O2/-Os to -O3. However, when using -Ofast there is a performance regression compared to -O3.

The Smallpt workload improves nicely as the compiler optimizations go harder.

Between no optimizations and the highest-level optimizations (-Ofast / -O3), the FLAC audio encoding performance is about five times faster.

As is the case when commonly running source-based benchmarks at Phoronix, the -O3 level generally does the best and is worth any extra added time to the compilation process. While the SLP Vectorizer wasn't covered in this article, it's for the -O3 level where LLVM developers are now looking at enabling it by default (among other potential optimization levels). The official release of LLVM 3.4 is likely to come around the end of 2013.