RFR: 8215560: [lworld] Investigate if method handle calls can be further optimized

[Roland Westrelin]

http://cr.openjdk.java.net/~roland/8215560/webrev.00/

Code patterns with a merge of 2 method handle invokes, one inlined and
of known return value and one not inlined of unknown return value used
to be optimized before Lworld. They are harder (if at all possible) to
optimize in Lworld and that use case seems a lot less important now. So
I chose to remove IR verification for those tests.

Other than that I found that return with the new calling convention
would sometimes keep some useless value buffering alive. I added a pass
to find those. Also, I noticed that sometimes, with incremental
inlining, some useless CheckCastPP would be left in the IR. I added a
point fix for that.

Roland.