Hi Dean,
On 10/09/2018 5:31 PM, [email protected] wrote:
Hi David. If a class references its own fields or methods, and that
code is executed by the interpreter, then we should expect that constant
pool entry to be resolved, shouldn't we?
Yes
Likewise the compiler will
treat it as resolved. If we treat is as unresolved, we run into the
case where we executed compiled code for a method that would have
resolved the constant pool entry, but JVMTI says we didn't.
I don't think the this_klass reference necessarily falls under that
categorization though. In an empty class (as used in the test case)
there's no bytecode that references any CP index for the current class.
It's simply an artifact of the classfile format.
Isn't the test in error for assuming no eager resolution?
Hmmm. Yes I suppose that is right. If we did eager resolution then even
the this_klass CP entry would have been resolved and the test would have
then encountered it from day one.
What bothers me is the self-referential nature of this ... and the
obscurity of it. I don't think anyone looking at the various heap
reference API's would think about a reference-to-self.
By the way, it looks like we always eagerly resolve this_class for
unsafe-anonymous classes.
Can you point me at that code please?
Thanks,
David
dl
On 9/9/18 10:51 PM, David Holmes wrote:
Bug: https://bugs.openjdk.java.net/browse/JDK-8210512
Webrev: http://cr.openjdk.java.net/~dholmes/8210512/webrev/
After the fix for JDK-8209361 where we modified JVM-TI to treat an
unresolved CP klass entry to a loaded klass as a resolved CP entry,
the listed test starting failing due to finding an extra reference to
the test class. As outlined in the bug report this extra reference:
17: instance of java.lang.Class(reflected
class=nsk.share.jdi.TestClass1, id=792)
actually comes from the class itself. Every classfile has a
self-referential entry in the constant pool (this_klass in JVMS 4.1)
and that is what we were encountering here.
I would argue that such an unresolved reference from a class to itself
should never be treated as a "real" reference from a JVM TI
perspective, and so we ship skip it - which is what the fix does:
- if (klass == NULL) {
+ if (klass == NULL || klass == ik) {
continue;
Testing: the test concerned
joatc testing of test that failed in 8209361 (TBD)
mach5 tiers 1-3 (TBD)
jvmti (TBD)
As noted testing is still TBD other then actual test but there are
technical delays with that so I'll get the RFR out anyway.
Thanks,
David
