Himeno didn't see any major change in performance between the four tested compilers.
The multi-threaded C-Ray ray-tracing benchmark did see a significant performance improvement when moving from LLVM/Clang 3.1 to LLVM/Clang 3.2, which now puts its performance in line with GCC 4.7.2. However, GCC 4.8.0 is already carrying additional optimizations that offers greater performance than GCC 4.7.2 / Clang 3.2.
Smallpt does poorly on Clang 3.2 due to lacking OpenMP support.
There isn't much difference between the compiler four-way for the MP3 and FFmpeg encoding benchmarks.
LLVM/Clang 3.2 is faster than GCC for the Tachyon ray-tracer.
Overall, LLVM/Clang 3.2 performed quite well against GCC 4.7/4.8. The GNU Compiler Collection had its share of wins in a number of the computational benchmarks, but LLVM/Clang 3.2 also had several wins. In cases where Clang wasn't the winner, in a majority of the cases was at least the open-source compiler was generally competitive aside from the notable lack of OpenMP support. Beyond the performance of the generated binaries, each compiler has its own set of features and abilities such as debugging and error messages, tuning switches, and other differences that have the potential to impact both developers and end-users of the compiled binaries.
Look for more benchmarks on the way.