We already have code in InlineTypeNode::make_scalar_in_safepoints that is supposed to prevent keeping field loads live just for safepoint debug info when we could instead use a "free" oop (i.e., not keeping a buffer allocation live).

We should investigate if we can make that code more aggressive:
- Should we include ProjNodes?
- Should the code that steps through Phis also handle LoadNodes?
- Do we even need the recursive handling of Phis? We should add tests.

I did a "quick" experiment by adding LoadNodes to the "else if (!(in->is_Con() || in->is_Parm()))" check but that leads to test failures. I think the problem is that a load in one branch that now stays alive leads to a inline type buffering staying alive in another branch because there is a merge point (the phi). Not using the oop would not keep that buffering alive because the field values would have been merged instead. I think this is only a problem when stepping through Phis because otherwise, if we don't use the oop load, we would load the fields from that same oop and that would also keep the buffering live.