Loop strip mining to decreases safepoint poll frequency

[Roland Westrelin]

> The patch looks ok from afar ;-) A lot of this C2 wrestling looks like
> voodoo magic... I trust that you have tested it etc.

The biggest risk I think is that some optimization stops working
silently because of the extra CFG nodes around the counted loops: I
wrote a test case for each optimization I thought was affected and
verified by hands that it still triggers. I also ran specjvm 2008 and
looked at generated code for cases where performance drops.

> What's the plan for upstreaming this? Let it bake in Shenandoah for a
> while and then go upstream?

Something like that. Ideally it would become the default upstream as
well. What I'm seeing is that if the inner loop runs for too few
iterations, strip mining can cause performance regressions so I'm not
too optimistic about this being the default.

Roland.