Covariant overrides on the Buffer Hierachy

[David M. Lloyd]

On 04/22/2014 09:53 AM, Peter Levart wrote:
> On 04/22/2014 02:48 PM, David M. Lloyd wrote:
>> On 04/22/2014 04:17 AM, Peter Levart wrote:
>>> On 04/22/2014 12:02 AM, David M. Lloyd wrote:
>>>> Um, do we *know* that there is a performance cost to covariantly
>>>> overriding these methods?  There would seem to be enough information
>>>> to keep them monomorphic, if the optimizer is smart enough to inline
>>>> the bridge methods and the delegating override method.  The overridden
>>>> methods in addition can be final, meaning that in the 99% case that
>>>> you're invoking directly on the desired buffer type, it should be just
>>>> as optimizable for the same reason that the original methods were
>>>> optimizable.  The only potentially "slow" invocation path is if you
>>>> call the method on a Buffer reference, and even then it seems like
>>>> there's enough information to avoid slowness - and if not, then that
>>>> seems like a HotSpot problem that is very solvable ("if all overrides
>>>> of this method call super.xxx(), inline & eliminate them").
>>>
>>> It's more complicated than that. Maybe we need an expert for hotspot JIT
>>> to answer this question, but as the code is written in the Rickard's
>>> webrev, then the reasoning behind the JIT to keep the monomorphic
>>> dispatch would have to be more involving. Richard is doing the following
>>> (in ByteBuffer):
>>>
>>> @Override
>>> public ByteBuffer position(int newPosition) {
>>>      super.position(newPosition);
>>>      return this;
>>> }
>>>
>>> javac compiles each of the covariant overrides as two methods - one that
>>> actually "overrides" the virtual method in superclass (has the same
>>> signature) and calls the covariant-returning method with a virtual
>>> dispatch. So ByteBuffer.position(int) is compiled as:
>>>
>>> public ByteBuffer position(int newPosition) {
>>>      super.position(newPosition);
>>>      return this;
>>> }
>>>
>>> public Buffer position(int newPosition) {
>>>      // this is an invokevirtual for position:(I)Ljava/nio/ByteBuffer;
>>>      return (ByteBuffer) position( (int) newPosition);
>>> }
>>
>> If the methods were final, AFAICT it'd be more like this:
>>
>> public final ByteBuffer position(int newPosition) {
>>     // iirc super upcall is already bytecoded as invokespecial
>>     invokespecial (Buffer)Buffer.position(newPosition);
>>     return this;
>> }
>>
>> public synthetic final Buffer position(int newPosition) {
>>     return effectively-invokespecial
>> (ByteBuffer)ByteBuffer.position(newPosition);
>> }
>>
>> Since there would only be one possible target for the invokevirtual,
>> my understanding is that the JIT will convert that into an
>> invokespecial, letting the whole works be optimized at worst and
>> inlined at best.
>>
>
> There are multiple possible targets for invokevirtual
> position:(I)Ljava/nio/Buffer; - all the methods that override it in all
> subclasses loaded. It doesn't matter if they are final or not (only if
> they are effectively final or not). The non-finality of a method has a
> performance impact only if the method *is* overridden in any of the
> loaded subclasses, otherwise it is effectively final and treated as such
> by JIT (at least that's how I understand it - any hotspot JIT expert
> please debunk my claims).

That is only true if you are calling the method on the base class, which 
is normally seldom done (though it will be nearly always in legacy code, 
but see below).

> That might also be the answer to why the synthetic method need not be
> marked as final if the covariant method is. The synthetic method can
> never be overridden in a sub-class (at least not by javac) - only the
> covariant method can be.

Doesn't sound quite right to me, but I'll defer to any experts who might 
wish to discuss it on compiler-dev.

> But as Paul noted, the methods on Buffer are probably not used in hot
> loops alone, since they are just for reading/adjusting
> position/limit/mark. The hot loops probably also contain methods for
> reading/writing the buffer and those are only defined on particular
> sub-types of java.nio.Buffer, so it can reasonably be expected that the
> static (guaranteed) type of target upon which methods are called in hot
> loops is a particular subtype of java.nio.Buffer and JIT only has one
> method to choose from in this case.

Yeah new code will call the covariant final method directly in almost 
all cases for sure, having been compiled against it.  The tricky case is 
old code - even if the JIT can figure out that it is really a subclass 
being invoked upon, it'd still be forced to target the non-covariant 
synthetic method due to the specific method reference in the bytecode. 
Hopefully at this point though, as you say, the lack of overriding 
classes will be enough to optimize the dispatch.

-- 
- DML