The new logic for sharing adapters will smear together arguments of the same size without regard to whether they are oops or not. This is incorrect with the tagged stack interpreter since it must also set the tag bigs for them. We need a more flexible strategy.
This bug affects the following VM/NSK test:

    nsk/jvmti/AddCapabilities/addcaps003
Could you please add the complete list of vm options too next time to make the bug easier to verify?
It looks like the test takes care of the options, but I'm not able to reproduce the failure on /net/sqenfs-1.sfbay/export1/comp/vm/jdk/hsx/17/pit/b09/jdk7b83/product or on b08/jdk782. Tried to look in nightlies for it's failure, but no luck, only one strange linux failure which doesn't reproduce. I'm trying to verify it during b09 pit which is strange, since it's said to be committed in b10.
I tried to reporoduce it via
/net/sqenfs-1.sfbay/export1/tools/gtee/harness/ute/ute -tlists empty -jdk /net/sqenfs-1.sfbay/export1/comp/vm/jdk/hsx/17/pit/b09/jdk7b83/product/solaris-sparcv9 -test nsk/jvmti/AddCapabilities/addcaps003 -vmopts "-Xmixed -XX:+UseConcMarkSweepGC -XX:+CMSIncrementalMode"
and also on solaris-amd64 and on linux-amd64