John, David, thanks for the feedback!
What do you think about the following version:
http://cr.openjdk.java.net/~vlivanov/8057919/webrev.02/jdk
http://cr.openjdk.java.net/~vlivanov/8057919/webrev.02/hotspot
Best regards,
Vladimir Ivanov
On 4/10/15 10:15 PM, John Rose wrote:
On Apr 9, 2015, at 4:16 PM, David Holmes <[email protected]
<mailto:[email protected]>> wrote:
Hi Vladimir,
On 10/04/2015 12:51 AM, Vladimir Ivanov wrote:
David, thanks for the feedback!
Updated webrev:
http://cr.openjdk.java.net/~vlivanov/8057919/webrev.01/jdk
http://cr.openjdk.java.net/~vlivanov/8057919/webrev.01/hotspot
I restored original logic and call into VM only if it fails.
This change seems fine to me, but I think John may prefer to see
getSimpleName implemented such that it always returns the name from
the innerclass attribute. (Though perhaps with caching if performance
is a concern?)
Yes, I do prefer the new logic instead of a combination of old and new.
Two concerns: Somebody somewhere might be relying on a bug in the old
logic and be frustrated by seeing that bug fixed. This is a fine line
to walk, but in this case I think the behavior change (if any) will show
up in places where we already have outages. Perhaps someone is
post-processing the result of getSimpleName to correct it. If so they
may no longer need to do that.
Second, the new logic may itself have issues. A typical problem with
the InnerClasses attribute is it is broken or stripped, or a related
nestmate is missing. (See parallel thread in core-libs-dev on 8076264.)
But the proposed change takes effect after a call to
Class.getEnclosingClass, which uses
InstanceKlass::compute_enclosing_class_impl and accesses the same
InnerClasses record.
This leads me to a comment: The common code (two uses
of InnerClassesIterator with klass_name_at_matches) should be factored
out. The factored subroutine should search out the self-record in the
InnerClasses attribute. The logic is tricky enough that I'm
uncomfortable with it being duplicated. Key issue: We don't want to
load any classes doing this lookup.
(For bonus points, factor the third use in that file, the one that looks
not for self but for children-of-self. Add a boolean flag that varies
the test.)
Bottom line: The new logic should replace the old.
The logic to compute simple name (Class.getSimpleName()) for
inner/nested/local classes is tightly coupled with Java naming scheme
and sometimes fails for classes generated from non-Java code.
Meta-question: if this is non-Java code then what does/should
"simpleName" even mean? "Returns the simple name of the underlying class
as given in the source code." If there is no (java) source code does
this have any meaning? Should it return "" ?
My reading of the JVMS is that InnerClasses attribute can be freely used
by non-Java compilers to store simple names for classes they generate.
The current wording for inner_name_index doesn't mention Java language.
True, but it does refer to source code: "represents the original
simple name of C, as given in the source code from which this class
file was compiled. " which seems misplaced as we're discussing classes
for which there is no source code as such.
It could be non-Java source code. In any case, the indicated component
of the "binary name" of the class is well-defined, whether or not it
occurs in source code. Note also that classes might be generated by ASM
(no source code per se) but we still have a reasonable expectation of
reflecting on them, even though reflection is (mostly) defined in terms
of source code constructs.
Anyway it just flags to me that perhaps these Class methods need to be
generalized a bit given the support for non-Java languages on the JVM.
But that's for core-libs folk to decide.
Yes, I think that's the issue. The multi-language folks (like me) would
be disappointed to hear that we decided that non-Java languages don't
have any needs here; they do.
Thanks,
— John
Thanks,
David
Best regards,
Vladimir Ivanov
Instead of parsing class name and try to extract simple name based on
JLS rules, the fix I propose is to use InnerClasses attribute from the
class file. Simple name is already recorded there.
>
JVMS-4.7.6: The InnerClasses Attribute
"inner_name_index: If C is anonymous (JLS §15.9.5), the value of the
inner_name_index item must be zero. Otherwise, the value of the
inner_name_index item must be a valid index into the constant_pool
table, and the entry at that index must be a CONSTANT_Utf8_info
structure (§4.4.7) that represents the original simple name of C, as
given in the source code from which this class file was compiled."
Since I consider backporting the fix into 8u60, I'd like to hear
opinions about backward compatibility of such change.
As an alternative solution, I can restore original logic and consult
InnerClasses attribute when class name parsing logic fails.
I think I'd prefer to see the VM call only used as a fallback if the
regular parsing fails. That would prevent any compatibility issues for
the 8u backport (ignoring tests that deliberately generate invalidly
named classes).
Thanks,
David
Testing: regression test, jck-runtime/java_lang, jdk/test/java/lang/
Thanks!
Best regards,
Vladimir Ivanov
