Loop strip mining to decreases safepoint poll frequency

[Roland Westrelin]

> I would have thought it would be useful in uncounted loops too.  I

Yes, I suppose it could be made to work with uncounted loops. Any
evidence that it would pay off would be nice...

> suppose there's some risk that adding the extra bookkeeping might
> slow down such loops.  But I was thinking that it might even allow
> vectorization in loops where it can't be done today.

I'm not sure I see how that would work. Applying strip mining on
uncounted loops would turn them into counted loops because we would need
to count the number of times we go through the loop to exit every few
iterations but the induction variable wouldn't be used by the code in
the loop for anything else. I would also have expected that programmers
naturally write counted loops that c2 would recognize as such and that
what was preventing vectorization from happening was something else.

Roland.