A number of people have suggested that GCC is the problem. If so, perhaps we should be testing with Intel's C compiler - http://www.intel.com/cd/software/products/asmo-na/eng/compilers/clin/index.htm - to see if this offers an improvement. There is an evaluation available for commercial use, and a free download for non-commercial (which by Intel's definitions means only developers who are not being paid) use.