[hibernate-dev] Bytecode interception and attribute name versus attribute index

[Sanne Grinovero]

Sounds like a very interesting patch, I agree with you that there's
much potential.
Although Steve from your description it seems like you don't think it can work?

I would expect something among the lines of what Gunnar described to
be feasible: to let the persisters figure out the right numbers when
they are created (once only). But I am not familiar enough with this
code area to get in the details.

About the method you mentioned " $$_hibernate_read_name()" : who
invokes that, and when is the code which invokes that method
generated?
I'm assuming that would be part of the persister, so it means
persister implementations are defined at runtime via bytecode
manipulation?



On 12 November 2015 at 15:18, Gunnar Morling <gunnar at hibernate.org> wrote:
>> You want us to delay enhancement until after the persisters are built.
>
> No, I mean to let enhancement define the attribute ordering and
> initialize persisters with that previously defined order.
>
>> What I suggest longer term is a slight variation...
>
> Yes, that sounds nice.
>
>
> 2015-11-12 16:07 GMT+01:00 Steve Ebersole <steve at hibernate.org>:
>> But see that is exactly the problem.  That cannot work.  You want us to
>> delay enhancement until after the persisters are built.  But building the
>> persisters requires access to the Class.  But by that time since we have the
>> Class reference we can no longer enhance it.  It's a chicken-egg problem.
>>
>> What I suggest longer term is a slight variation, where we drive the
>> "indexing" based on the mapping model.  Actually that also will not work as
>> the code is today.  Really what I'd like to do is to leverage the
>> jandex-binding work.  Specifically there we have a Jandex-driven pre-parsing
>> of the domain model (including XML augmentation).  That model would be
>> perfect for this.
>>
>> On Thu, Nov 12, 2015 at 9:00 AM Gunnar Morling <gunnar at hibernate.org> wrote:
>>>
>>> It would be a one-time effort during persister initialisation; At that
>>> time a look-up would happen in order to seed persisters with
>>> attributes in the order as determined by the enhancer. Then at
>>> runtime, upon getter/setter invocation, the index values passed from
>>> the enhanced code would be used as is for array index access in
>>> persisters.
>>>
>>> 2015-11-12 15:30 GMT+01:00 Steve Ebersole <steve at hibernate.org>:
>>> > But I think that just redefines the same problem.  Just that now instead
>>> > of
>>> > String->Integer we need some Integer->Integer resolution.  Not sure that
>>> > really gains us anything.
>>> >
>>> > On Thu, Nov 12, 2015 at 8:29 AM Steve Ebersole <steve at hibernate.org>
>>> > wrote:
>>> >>
>>> >> I have no estimate as to how much this would help.  Many of the
>>> >> improvements coming from the performance work have been of the sort
>>> >> that is
>>> >> removing Map lookups; some of them across VERY small (and finite) sets
>>> >> of
>>> >> Map keys.  My concern here is that the Map look up occurs any time an
>>> >> enhanced entity's getter or setter is called, which seems like a
>>> >> potentially
>>> >> bigger performance impact.
>>> >>
>>> >> I'm going to table this for now until the jandex-binding work.  But as
>>> >> part of that work, I think it makes a ton of sense to make sure we
>>> >> develop a
>>> >> strategy for consistent ordering of processing attributes.
>>> >>
>>> >> On Thu, Nov 12, 2015 at 2:36 AM Gunnar Morling <gunnar at hibernate.org>
>>> >> wrote:
>>> >>>
>>> >>> Is there any estimate what would be gained perf-wise by doing this
>>> >>> optimization?
>>> >>>
>>> >>> > First, this obviously is more fragile than relying on names; if the
>>> >>> > order
>>> >>> > changes but the bytecode is not re-enhanced, that could lead to very
>>> >>> > subtle, nasty problems.
>>> >>>
>>> >>> Agreed, but to me it seems acceptable with such optimization having to
>>> >>> be explicitly enabled and the docs of that setting describing this
>>> >>> matter clearly.
>>> >>>
>>> >>> > The other issue is that this requires us to be able to consistently
>>> >>> > be
>>> >>> > able
>>> >>> > to order the attributes.
>>> >>>
>>> >>> Maybe the enhancer could enhance entities with a (static) method
>>> >>> returning the attribute order it works with, or maybe some separate
>>> >>> contract for making this order available. Then at runtime this order
>>> >>> could be fed to persisters et al. for enhanced entities.
>>> >>>
>>> >>> Not sure whether it's doable or how complex it'd be, just throwing out
>>> >>> the idea.
>>> >>>
>>> >>> --Gunnar
>>> >>>
>>> >>>
>>> >>> 2015-11-11 21:37 GMT+01:00 Steve Ebersole <steve at hibernate.org>:
>>> >>> > As I work on HHH-10267, a thought that keeps coming up in my head is
>>> >>> > that
>>> >>> > it would be great to avoid Map lookups in the enhancement and
>>> >>> > interception
>>> >>> > code.
>>> >>> >
>>> >>> > As an example of what I mean, consider the reader for a String field
>>> >>> > named
>>> >>> > `name` which roughly looks like:
>>> >>> >
>>> >>> > public String $$_hibernate_read_name() {
>>> >>> >     return (String) $$_hibernate_getInterceptor().readObject( this,
>>> >>> > "name",
>>> >>> > this.name );
>>> >>> > }
>>> >>> >
>>> >>> > In the interception code we need to resolve the attribute name
>>> >>> > ("name"
>>> >>> > above) to the pertinent information from the persister,etc.
>>> >>> > Typically
>>> >>> > this
>>> >>> > is a Map lookup; sometimes it is a done by iterating a String[] of
>>> >>> > the
>>> >>> > attribute names.
>>> >>> >
>>> >>> > Ultimately the information is taken from persister and friends,
>>> >>> > which
>>> >>> > means
>>> >>> > that the information is generally already readily consumable via
>>> >>> > array
>>> >>> > access once we have resolved the name to a proper index.  Which got
>>> >>> > me
>>> >>> > thinking it would be great if we could encode the attribute index
>>> >>> > into
>>> >>> > the
>>> >>> > enhanced entity directly.  Going back to the example about.. if we
>>> >>> > knew
>>> >>> > that the "name" attribute was the 4th attribute (index=3) according
>>> >>> > to
>>> >>> > the
>>> >>> > entity persister we could leverage that information for better
>>> >>> > performing
>>> >>> > access to the attribute metadata in the enhanced code and
>>> >>> > interceptor:
>>> >>> >
>>> >>> > public String $$_hibernate_read_name() {
>>> >>> >     return (String) $$_hibernate_getInterceptor().readObject( this,
>>> >>> > 3,
>>> >>> > this.name );
>>> >>> > }
>>> >>> >
>>> >>> > One gotcha - always has to be one devil right ;)  Ok, 2 really...
>>> >>> >
>>> >>> > First, this obviously is more fragile than relying on names; if the
>>> >>> > order
>>> >>> > changes but the bytecode is not re-enhanced, that could lead to very
>>> >>> > subtle, nasty problems.
>>> >>> >
>>> >>> > The other issue is that this requires us to be able to consistently
>>> >>> > be
>>> >>> > able
>>> >>> > to order the attributes.  The Enhancer currently does not rely on
>>> >>> > the
>>> >>> > built
>>> >>> > metadata (org.hibernate.mapping) at all; it parses the entity
>>> >>> > annotations
>>> >>> > (completely annotation specific) on its own.  Given the 2-phase
>>> >>> > break
>>> >>> > down
>>> >>> > in JPA bootstrapping, having Enhancer leverage the built metadata is
>>> >>> > not
>>> >>> > really going to be possible.  Which is unfortunate, because doing so
>>> >>> > would
>>> >>> > be nice for other reasons (like supporting XML mappings for
>>> >>> > enhancement
>>> >>> > as
>>> >>> > well, etc).
>>> >>> >
>>> >>> > That's a lot of information, sorry about that :)
>>> >>> >
>>> >>> > But anyone have thoughts on that?
>>> >>> > _______________________________________________
>>> >>> > hibernate-dev mailing list
>>> >>> > hibernate-dev at lists.jboss.org
>>> >>> > https://lists.jboss.org/mailman/listinfo/hibernate-dev
> _______________________________________________
> hibernate-dev mailing list
> hibernate-dev at lists.jboss.org
> https://lists.jboss.org/mailman/listinfo/hibernate-dev