[gtani]

and this phrase on page 7 highlighted

    The benchmarking itself was done in a simple and 
    fairly un-scientific fashion.

    ... older Pentium IV workstation. Run-times were
    measured using wall-clock time.

but still a good read about tweaking c++ (I think at the time, clang and gcc were still implementing the C++11 standard features).

Also i remember long HN threads about this in mid 2011 but i can't find them, here'a a later discussion https://news.ycombinator.com/item?id=4539668