However, GCC 4.7's just around the corner and I haven't tried that yet...
Though, admittedly, if speed is an issue you probably have already manually loop-unrolled and used the gcc compiler intrinsics.
A) it's proprietary, I have no interest in relying on a proprietary toolchain (and from what I gather neither does the company I work for)
B) it supports a very limited range of cpu architectures, not only is it directly tailored for Intel cpu's it even has a history of selecting poor code paths for AMD cpu's.
As for manual loop unrolling, for alot of code PGO does a great job here by unrolling based on the statistics gathered during the first pass. In fact GCC's pgo seems to do a better job than ICC's pgo implementation, ICC's lto beats GCC's on the other hand, and of course ICC does a better job at vectorization and has better optimized math functions.
That said I use both, and at work we test our code against both toolchains (and some other compilers aswell). The static analyser in Clang is a welcome addition and the error diagnostics/reporting is top notch so it certainly has strong features even though it falls behind GCC in code optimization.