> How can something, which is faster on slower machines (than what?)
> be slower on faster machines?

[this is an aside..]

As a generic example, take "Duff's device" for loop unrolling.

It's faster on slower machines, because it eliminates loop tests.

It's slower on modern, fast, superscalar machines because the "branch
into the middle" aspect interferes with instruction scheduling and
software pipelining optimizations *and* it's got a bigger icache
footprint because the body is N times larger.

					- Bill