Draft JEP Announcement: "Computed Constants"

[forax at univ-mlv.fr]

----- Original Message -----
> From: "John Rose" <john.r.rose at oracle.com>
> To: "Remi Forax" <forax at univ-mlv.fr>
> Cc: "Dan Heidinga" <heidinga at redhat.com>, "Per-Ake Minborg" <per-ake.minborg at oracle.com>, "leyden-dev"
> <leyden-dev at openjdk.org>
> Sent: Saturday, July 29, 2023 1:58:49 AM
> Subject: Re: Draft JEP Announcement: "Computed Constants"

> On 28 Jul 2023, at 13:53, Remi Forax wrote:
> 
>> I'm fine with a method <class-field-init> that takes an int (the index from the
>> AutonomousValue attribute) and returns an Object to initialize each autonomous
>> field at the first call. For me, there is no need for a BSM here.
>>
>> Also I prefer the CC initialization semantics (the field initializer is only
>> executed once) to the one you are using in your prototype, the field
>> initializer can be called several times concurrently but only store the first
>> value that complete.
> 
> Yeah, after polishing my prototype it might end up somewhere like that.
> 
> The benefit of the CC stuff is that it can clearly provide and document the
> correct high-level semantics.  

yes !

>. And the low-level VM machinery can be simpler
> (fewer guarantees), right?  Or we could do an all-in-one solution where the
> JVMS promises to mimic the CC logic example.
> 
> But I prefer a layered design, where the VM uses a BSM to route through the CC
> logic, which then elevates the result to what the JLS or the CC API will
> guarantee.  So the VM can use the condy-like binding it is accustomed to, which
> uses CC directly to get the right mutex behavior.

I do not think a layered design will be the end game, paying the transition through a BSM for each initialization may be more expensive that the initialization of the constant itself. That's one of the challenge of the computed constants, we know the initialization is done only once, so the VM has to be more than with invokedynamic / constant dynamic. 
So in the end, my gut feeling is that the VM will have to do the job, like it does currently with a <clinit>.

But the layered design / using a BSM is a nice middle step, to help us to find the right semantics.

Rémi