Will lambdas be inline?

Remi Forax forax at univ-mlv.fr
Sun Apr 12 17:52:44 UTC 2020


----- Mail original -----
> De: "Brian Goetz" <brian.goetz at oracle.com>
> À: "valhalla-dev" <valhalla-dev at openjdk.java.net>
> Cc: stevenselectronicmail at gmail.com
> Envoyé: Dimanche 12 Avril 2020 18:15:29
> Objet: Fwd: Will lambdas be inline?

> The below was received on the -comments list.
> 
> Short answer: Yes, this is very much under consideration; the identity
> of a lambda is irrelevant, and JLS 15.27.4 makes this clear:
> 
>>     This implies that the identity of the result of evaluating a
>> lambda expression (or, of serializing and deserializing a lambda
>> expression) is unpredictable, and therefore identity-sensitive
>> operations (such as reference equality (§15.21.3), object locking
>> (§14.19), and the System.identityHashCode method) may produce
>> different results in different implementations of the Java programming
>> language, or even upon different lambda expression evaluations in the
>> same implementation.
>>
> 
> It is likely that this change will expose more opportunities for escape
> analysis, though it would likely not expose significant flatness/density
> optimizations since field descriptors / array components would never
> name the inline proxy directly.  Still, this is exactly the sort of
> optimization we had in mind when we penned the above flexibility in 15.27.4.
> 
> As to accidental behavioral compatibility concerns, this change will
> make _more_ lambdas equal to each other, since `==` on two lambda
> instances would no longer be an identity comparison, but a comparison of
> all the state of the lambda, so two otherwise-identical but separately
> constructed lambda proxies would be `==` if they were inline but might
> not be under the current situation, but it is unlikely to go in the
> other direction.
> 
> On the other hand, it will make `==` operations on lambdas slower.

I wanted to test the perf again, but it seems that currently == is broken :(

If i run
https://github.com/forax/valuetype-lworld/blob/master/src/main/java/fr.umlv.valuetype/fr/umlv/valuetype/StackOverflowACMP.java

It prints true.


> 
> Overall I'd like to see this change made, but we have to work through
> the details to ensure that it is not excessively surprising.

Rémi

> 
> 
> -------- Forwarded Message --------
> Subject: 	Will lambdas be inline?
> Date: 	Sat, 11 Apr 2020 21:14:22 -0700
> From: 	Steven Stewart-Gallus <stevenselectronicmail at gmail.com>
> To: 	valhalla-spec-comments at openjdk.java.net
> 
> 
> 
> Hi
> 
> I was curious if lambdas will be inline?
> I feel this is pretty natural.
> You're already not really supposed to be doing these kinds of things with
> lambdas.
> Technically you could spin both depending on the situation.
> 
> Correct me if I'm wrong but under the current plan you'd need to use
> InlineCont for the best performance.
> 
> interface Cont {
> Cont next();
> }
> 
> interface InlineCont extends InlineObject {
> InlineCont next();
> }
> 
> interface IdentityCont extends IdentityObject {
> IdentityCont next();
> }
> 
> Slightly unfortunate because I think it's highly unlikely many interfaces
> will break compat for a slight perf boost.
> 
> And if lambdas are inline than would IdentityCont not be a functional
> interface?
> Thanks
> Steven Stewart-Gallus


More information about the valhalla-dev mailing list