RFR: several small changes including 8207040: [lworld] TestLWorld fails with "Meet Not Symmetric" with -Xcomp

[Roland Westrelin]

http://cr.openjdk.java.net/~roland/8207040/webrev.00/

This includes the fix for the type system bug 8207040 and a test case
that triggers it. The change mimics what happens with regular instances.

In the process, I found that:

- checkcast buffers values when it's not needed. The change delays
  allocation. I actually wonder if the path where we allocate is seen in
  practice and if it is if we need to allocate at all.

- the Parse::merge_common sometimes buffers a value when there's no
  merge point which is unnecessary so I changed that logic too.

With both changes, ValueTypeNodes are seen in a few places where they
are not expected. This change fixes those. In particular an array store
can see a ValueTypeNode when storing to a non flattened array. The
current code treats it as a flattened store because its decision are
based on the type of the value being stored (ValueTypeNode or not). I
reworked that logic.

I'm a bit confused by monitorenter. We deoptimize when we see a value
but we're already at the monitorenter. So when we deoptimize, given we
don't reexecute the monitorenter, we resume execution after that
bytecode. How does that trigger an exception?

Roland.