In the fix for 7009361 I added some logic in deoptimization to adjust interpreted callers by the size of the signature at the call site instead of the size of the callee since they can be different in the presence of invokedynamic. In our current system things functioned ok without because we tends to leave a lot of extra space on the stack. There was a bug in the code I added with reduced the value to either 0 or 1 which caused problems in rare cases. The fix is to correct the grouping to return the proper answer.
The failing test was:
Could you please specify full test name next time, it would really help bug verification process. Link to run results could also be helpful to find out vm options to use for reproducing. Although, may be 64bit mode is enough.