The image sharpening operation within GraphicsMagick improved by 60% with GCC 4.7.0 when it was built with the -march=bdver1 compiler flag over -march=k8. The Open64 bdver1 optimization resulted in a 25% improvement.
When running Himeno, LLVM/Clang with just the "-march=k8" flag did the best but regressed sharply when using the bdver1-level optimizations. Meanwhile, GCC 4.7.0 improved by 12% with bdver1 support and Open64 was up by 5% with this compiler tuning.
When running HMMer, Open64 4.5.1 produced the fastest binaries ahead of both GCC 4.7.0 and LLVM/Clang 3.1 SVN. The bdver1 support provided minimal benefit in this test.