what tests have been done to confirm this?
Lots of different tests. Jim Brown wrote some Sieve of Eratosthenes code in different basic languages and I wrote the C++ reference (that is using no optimizations, just a straight port that does the same ammount of work) code. There's a thread about it somewhere complete with links to sourcecode.Also a series of tests involving various Krushkal and Dijkstra implementations, and large scale (4GB) duplicate integer finder algorithm, showed the same consistent 80% performance.
hmm, one thing about C++, different compilers get different results.
The tests I did was compiled with the same compiler (the MinGW version of GCC) that powers BlitzMAX for most consistently useful and comparable data.