Hi Paul,
On 23/05/2018 2:05 AM, Paul Sandoz wrote:
Hi David,
Thank you for your patience, i struggled to explain my point.
How about we proceed as is, and as you suggest, i can discuss more with John
when he is back from vacation. I think we will have time to revisit if
necessary.
That sounds good to me. As I noted this would not involve any spec
change so there is no concern with making an implementation change later.
My preference would be to store the classes from the class file bytes in a Set,
cache in ReflectionData, then obtain the “cloned” array from that cached Set.
We can use Set.of, which is efficient for small sizes, for the case of one
element we could even erase the cached value to Object.
The only glitch there is you then need to find the nest-host and swap it
into the zeroth element.
Thanks,
David
Thanks,
Paul.
On May 21, 2018, at 10:50 PM, David Holmes <david.hol...@oracle.com> wrote:
Hi Paul,
On 22/05/2018 2:15 PM, Paul Sandoz wrote:
Hi David,
On May 21, 2018, at 5:05 PM, David Holmes <david.hol...@oracle.com> wrote:
Hi Paul,
On 22/05/2018 2:39 AM, Paul Sandoz wrote:
On May 20, 2018, at 11:32 PM, David Holmes <david.hol...@oracle.com> wrote:
3984 public Class<?>[] getNestMembers() {
I still think not removing dups is a mistake as it could be a source of subtle
bugs. But i doubt at this point i can persuade you or others to change it :-)
Unlikely. :) Again well-formed programs just won't encounter this and it would
penalize all well-formed programs.
Although those well-formed programs may need to check for dups themselves
because they don’t want to rely in implementation details (and they are not
aware of the probability of implementations deviating).
I'm quite concerned about your level of concern with "dups". This just
shouldn't be an issue. While the spec allows for dups javac will never produce them - and
file a bug on it if it ever does! Similarly for other compilers - there is no reason to
generate duplicate entries.
Perhaps i am obsessing a little too much, i thought there might be a slight
window of opportunity while other related reviews are progressing :-) but i
don’t want to block things for 11.
My concern, placing my library/API designer hat on, is the specification is
saying something very clear and yet on the other hand it's as if we are saying
“oh you can ignore that, the specification does not matter, it will never
happen in practice”. It feels like the JVM world is intruding too much into the
reflection world (see below).
Looking through the JVMS and the defined classfile attributes it seems to me
that the annotations[] of RuntimeVisibleAnnotations (et al) doesn't preclude
duplicates either. And the bootstrap_methods[] of the
BootstrapMethodsAttribute. Also look at the parameters[] of the
MethodParametersAttribute**. Do you agree?
** Which even has an explicit note this is not something a JVM implementation
has to check.
I don’t dispute there may be duplicate information in class file bytes, nor am
i suggesting the class file bytes for nestmates be changed, nor that the
verifier get involved. However, the reflection API is not a direct reflection
of those class file bytes. It provides a runtime view of a class and often
performs its own computation from and validation of information in the class
file bytes.
Not all information in the class file bytes is directly accessible via the
reflection API, such as BootstrapMethodsAttribute, or
MethodParametersAttribute, the latter of which AFAICT is exposed via the
Reflection API indirectly via a Parameter[] array returned by
Method.getParameters().
I am less sure about RuntimeVisibleAnnotations but there is quite a lot of
processing performed by the Java reflection code before annotations reach the
hands of the developer, and a quick look at some code shows the use of maps
keyed by annotation class to the annotation value. And see for example here in
AnnotationParser:
if (AnnotationType.getInstance(klass).retention() == RetentionPolicy.RUNTIME
&&
result.put(klass, a) != null) {
throw new AnnotationFormatError(
"Duplicate annotation for class: "+klass+": " + a);
http://hg.openjdk.java.net/jdk/jdk/file/95ba3a1dc2b2/src/java.base/share/classes/sun/reflect/annotation/AnnotationParser.java#l110
Okay I acknowledge your point here about VM view versus Reflection view. It
would be easier if we were returning the members in a duplicate-detecting data
structure rather than an array (though placing the host at zeroth element
complicates that). The sheer effort involved in detecting and removing
duplicates from an array is what made me shy away from pushing for that. You
can see the EG discussion from here:
http://mail.openjdk.java.net/pipermail/valhalla-spec-experts/2017-December/000464.html
If you really think this is worth re-opening it would probably be expedient to
discuss with John so you both end up on the same page, then let me know the
outcome.
Though I'll also note that we can strengthen the current implementation at any
time and just update the @implNote.
Thanks,
David
-----
Here’s a thought: did you consider caching the nest members in the
ReflectionData class? that may be worth doing regardless of dups.
No that was not considered. Caching, as you know, is a space-time trade off and
we have no data to use to determine whether caching would be of any benefit
here. To me that would be a future RFE regardless. (And I don't expect these
introspective nest methods to be used much in any case.)
Yes, agreed, caching can be possible future work.
Paul,
