When testing the Apache web-server on the AMD Llano system, the performance between the GCC 4.6.1 and LLVM 2.9 Clang-produced binaries were close. For this core Linux server process there was not a definitive advantage to one compiler over the other at this time for Llano.
C-Ray, our favorite multi-threaded ray-tracing test, did much better with GCC than LLVM/Clang with the latest release versions of both. There were no improvements with GCC 4.6.1 over GCC 4.5.3.
When running POV-Ray, which is currently single-threaded ray-tracing, the binary produced by Clang was faster than GCC.