With C-Ray, one of the multi-threaded test favorites, using GCC 4.5.2 with DragonEgg and LLVM 2.9 produced the fastest binaries overall. The compiler difference for this multi-threaded ray-tracer under the Core i7 990X was not as huge as the slower systems, but the sweet spot did appear to be using GCC but with LLVM's optimizers and code-generators. Using Clang was also faster than GCC on the Intel Core i3 Sandy Bridge system, but slower on the two AMD computers and effectively the same with the Intel Core i7 990X and Intel Core 2 Duo E8400.
With Smallpt, another interesting test, the performance with Clang was horrible in comparison to the GCC compilers. In all tests, LLVM/Clang 2.9 was multiple times slower than GCC. It looks like Clang has issues with Smallpt when it comes to threading with Clang being twice as slow on the Intel Core 2 Duo, more than six times slower on the Intel Core i7 990X, nearly four times slower on the Quad-Core Opteron 2384, etc. In terms of comparing GCC 4.5.2, GCC 4.6.0, and DragonEgg, the performance was not that different between the three.