We are now around 6/7 times slower than the LLVM code.

There are some rounding errors in the tests, but otherwise
the implementation seems correct.

Also, fix for notifications throttle delay value

This reverts commit a60a1be7.