Scott writes:
> CM is not faster than NM, PERIOD. At best it could be equivalent.

Actually it's not inconceivable that in some case a CM task could be
faster than the same thing compiled in NM.  Keep in mind that NM code
and data takes up several times as much space as CM code.

It is possible that the combination of the CM Emulator and your CM code
and data might have better locality than the NM equivalent, which might
mean that it could all fit into cache at once, reducing the number of
cache and TLB misses to the point that it runs faster.

You might see this effect on a system like a 980 which has a large cache,
a *huge* cache-miss penalty, and a relatively dumb CPU when it comes to
prefetching and predicting branches etc.

Whether this has ever occurred in practice I have no idea, but it is an
interesting concept.  I first heard it suggested by Jim Miller of HP (the
primary author of OCT I believe) in 1988.

G.