I can't find any documentation nor can I find any information regarding it's implementation. I am asking this, as I focusing on fixing bugs that destroy currently has, most noticeably bugs regarding attributes. I am not sure that this requires a DIP in order to fix this.

https://issues.dlang.org/show_bug.cgi?id=17297
https://issues.dlang.org/show_bug.cgi?id=15246

On Friday, 19 January 2018 at 15:56:58 UTC, 12345swordy wrote:
> I can't find any documentation nor can I find any information regarding it's implementation. I am asking this, as I focusing on fixing bugs that destroy currently has, most noticeably bugs regarding attributes. I am not sure that this requires a DIP in order to fix this.
>
> https://issues.dlang.org/show_bug.cgi?id=17297
> https://issues.dlang.org/show_bug.cgi?id=15246

You'll probably learn new and interesting things but i warn you: this problem can't be fixed.

Look at here: https://github.com/dlang/druntime/blob/master/src/rt/lifetime.d#L1380.
rt_finalize calls rt_finalize2 which use the **dynamic** type info to get a pointer to the __xdtor of each generation.

On Friday, 19 January 2018 at 15:56:58 UTC, 12345swordy wrote:
> I can't find any documentation nor can I find any information regarding it's implementation. I am asking this, as I focusing on fixing bugs that destroy currently has, most noticeably bugs regarding attributes. I am not sure that this requires a DIP in order to fix this.
>
> https://issues.dlang.org/show_bug.cgi?id=17297
> https://issues.dlang.org/show_bug.cgi?id=15246

Have a look here: https://github.com/dlang/druntime/blob/master/src/rt/lifetime.d#L1380

On Friday, 19 January 2018 at 16:18:15 UTC, Basile B. wrote:
>this problem can't be fixed.
You need to specify, do you mean this can't be fixed without a DIP or even with a DIP?

On 1/19/18 10:56 AM, 12345swordy wrote:
> I can't find any documentation nor can I find any information regarding it's implementation. I am asking this, as I focusing on fixing bugs that destroy currently has, most noticeably bugs regarding attributes. I am not sure that this requires a DIP in order to fix this.
> 
> https://issues.dlang.org/show_bug.cgi?id=17297
> https://issues.dlang.org/show_bug.cgi?id=15246

So there has been very little definition around how destructors use and inherit attributes. Most of this is because destructors are generally called from the GC, where it doesn't matter at all what the attributes are.

I think before you can fix such things, we need a clear model of how the destructors are called, and what is inherited and what is overridden in base classes. Then we can think about how to fix the situation.

Note that:

a. The GC must have a *runtime* definition of how to call the destructors. And we need to statically disallow the GC from calling destructors it shouldn't be allowed to call (if such cases exist).
b. Object.~this can't be a decider of what the attributes should be, as this prevents adding any additional attributes.
c. There is a push recently (by Andrei and Lucia and others) to remove some of the "magic" calls from the compiler, and replace them with templates, so we can have more library control over what happens (and more comiple-time introspection). This should help quite a bit in the implementation of such things. But we definitely are limited by the virtual-ness of Objects, and the opaqueness of the GC.

-Steve

On Friday, 19 January 2018 at 18:22:33 UTC, Steven Schveighoffer wrote:
> On 1/19/18 10:56 AM, 12345swordy wrote:
>> I can't find any documentation nor can I find any information regarding it's implementation. I am asking this, as I focusing on fixing bugs that destroy currently has, most noticeably bugs regarding attributes. I am not sure that this requires a DIP in order to fix this.
>> 
>> https://issues.dlang.org/show_bug.cgi?id=17297
>> https://issues.dlang.org/show_bug.cgi?id=15246
>
> So there has been very little definition around how destructors use and inherit attributes. Most of this is because destructors are generally called from the GC, where it doesn't matter at all what the attributes are.
>
> I think before you can fix such things, we need a clear model of how the destructors are called, and what is inherited and what is overridden in base classes. Then we can think about how to fix the situation.
>
> Note that:
>
> a. The GC must have a *runtime* definition of how to call the destructors. And we need to statically disallow the GC from calling destructors it shouldn't be allowed to call (if such cases exist).
> b. Object.~this can't be a decider of what the attributes should be, as this prevents adding any additional attributes.
> c. There is a push recently (by Andrei and Lucia and others) to remove some of the "magic" calls from the compiler, and replace them with templates, so we can have more library control over what happens (and more comiple-time introspection). This should help quite a bit in the implementation of such things. But we definitely are limited by the virtual-ness of Objects, and the opaqueness of the GC.
>
> -Steve

I see, this does look like it needs a DIP in order to fix this. How the progress of the calls being replaced by templates by Andrei and company?

On Friday, 19 January 2018 at 17:01:50 UTC, 12345swordy wrote:
> On Friday, 19 January 2018 at 16:18:15 UTC, Basile B. wrote:
>>this problem can't be fixed.
> You need to specify, do you mean this can't be fixed without a DIP or even with a DIP?

I meant: at all. But i'd like to be wrong.

On 1/19/18 2:48 PM, 12345swordy wrote:
> I see, this does look like it needs a DIP in order to fix this. How the progress of the calls being replaced by templates by Andrei and company?

Here are the improvements that Lucia is doing/has done, which is a lot of good stuff:

https://github.com/dlang/druntime/pulls?utf8=%E2%9C%93&q=is%3Apr+author%3Asomzzz

But really, I think we need a DIP to clarify the model of destructors and attribute inheritance. Right now, there is nothing (just that you can have attributes): https://dlang.org/spec/class.html#destructors

There are some notes about destructors in the attributes section, e.g. "Pure destructors do not benefit of special elision." But there is nothing talking about how the attributes are inherited, or rules about when they can be called.

I would love for Object to have it's attribute shackles removed. It's not just destructors, but toString, opCmp, etc.

-Steve

On Friday, 19 January 2018 at 20:21:44 UTC, Steven Schveighoffer wrote:
> On 1/19/18 2:48 PM, 12345swordy wrote:
>> [...]
>
> Here are the improvements that Lucia is doing/has done, which is a lot of good stuff:
>
> [...]
I found out that this fix does require a DIP. (Thanks Walker!)