https://issues.dlang.org/show_bug.cgi?id=14363

ag0aep6g@gmail.com changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Keywords|                            |rejects-valid
                 CC|                            |ag0aep6g@gmail.com

--

https://issues.dlang.org/show_bug.cgi?id=14363

--- Comment #1 from Kenji Hara <k.hara.pg@gmail.com> ---
I'm not sure why SDC accepts the code.

The following reduced case is correctly rejected by dmd, because ba.foo() won't work expectedly. Maybe SDC generates wrong code?

void main()
{
    uint x = 7;

    class A
    {
        auto foo()
        {
            return x;
        }
    }

    auto makeB(uint y)
    {
        class B : A
        {
            auto bar()
            {
                return foo() + y;
            }
        }

        return new B(5);
    }

    A a = new A();
    assert(a.foo() == 7);

    auto b = makeB(3);
    assert(b.bar() == 10);

    A ba = b;
    assert(ba.foo() == 7);
}

--

https://issues.dlang.org/show_bug.cgi?id=14363

--- Comment #2 from Shammah Chancellor <shammah.chancellor@gmail.com> ---
Your code has a bug?

return new B(5);

Should be:

return new B();


With that, your example compiles and works just fine on SDC.


(In reply to Kenji Hara from comment #1)
> I'm not sure why SDC accepts the code.
> 
> The following reduced case is correctly rejected by dmd, because ba.foo() won't work expectedly. Maybe SDC generates wrong code?
> 
> void main()
> {
>     uint x = 7;
> 
>     class A
>     {
>         auto foo()
>         {
>             return x;
>         }
>     }
> 
>     auto makeB(uint y)
>     {
>         class B : A
>         {
>             auto bar()
>             {
>                 return foo() + y;
>             }
>         }
> 
>         return new B(5);
>     }
> 
>     A a = new A();
>     assert(a.foo() == 7);
> 
>     auto b = makeB(3);
>     assert(b.bar() == 10);
> 
>     A ba = b;
>     assert(ba.foo() == 7);
> }

--

https://issues.dlang.org/show_bug.cgi?id=14363

--- Comment #3 from Kenji Hara <k.hara.pg@gmail.com> ---
(In reply to Shammah Chancellor from comment #2)
> Your code has a bug?
> 
> return new B(5);
> 
> Should be:
> 
> return new B();

Yes. Sorry.

> With that, your example compiles and works just fine on SDC.

Hmm, interesting. SDC might use a thunk to get valid context pointer for base
classes from the instantiated context.
But it would need some runtime cost and additional vtbl entry. I don't have any
knowledge about the SDC backend, but the ABI around nested classes would be
different from dmd's.

--

https://issues.dlang.org/show_bug.cgi?id=14363

--- Comment #4 from Kenji Hara <k.hara.pg@gmail.com> ---
Maybe SDC inserts a hidden field to access enclosing scope in each derived classes?

void main() {
    uint x = 7;

    class A {
        int fieldA;
        auto foo() { return x; }
    }

    auto makeB(uint y) {
        class B : A {
            int fieldB;
            auto bar() { return foo() + y; }
        }

        return new B(5);
    }

    A a = new A();
    assert(a.foo() == 7);

    auto b = makeB(3);
    alias B = typeof(b);
    assert(b.bar() == 10);

    pragma(msg, A.fieldA.offsetof);
    pragma(msg, __traits(classInstanceSize, A));
    pragma(msg, B.fieldB.offsetof);
    pragma(msg, __traits(classInstanceSize, B));
}

If B.fieldB.offsetof + int.sizeof + (void*).sizeof <=
__traits(classInstanceSize, B), both B and A have their own hidden fields.

Note that it's intentionally disallowed by dmd. With dmd, implicitly insertion of hidden field is limited at most once. At best, it's an incompatibility between dmd and SDC.

--

https://issues.dlang.org/show_bug.cgi?id=14363

--- Comment #5 from Shammah Chancellor <shammah.chancellor@gmail.com> ---
I'm not sure about that part of SDC.  I will investigate and see if deadalnix has some input here.

--

https://issues.dlang.org/show_bug.cgi?id=14363

deadalnix <deadalnix@gmail.com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |deadalnix@gmail.com

--- Comment #6 from deadalnix <deadalnix@gmail.com> ---
(In reply to Shammah Chancellor from comment #5)
> I'm not sure about that part of SDC.  I will investigate and see if deadalnix has some input here.

SDC does it by sticking one context in the base and one context in the child.

--

https://issues.dlang.org/show_bug.cgi?id=14363

yebblies <yebblies@gmail.com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |yebblies@gmail.com
           Hardware|x86                         |All
            Version|unspecified                 |D2
            Summary|DMD should compile SDC      |[SDC] Error with inheriting
                   |test0165.d                  |nested classes in nested
                   |                            |functions
                 OS|Mac OS X                    |All

--

https://issues.dlang.org/show_bug.cgi?id=14363

--- Comment #7 from Kenji Hara <k.hara.pg@gmail.com> ---
(In reply to deadalnix from comment #6)
> (In reply to Shammah Chancellor from comment #5)
> > I'm not sure about that part of SDC.  I will investigate and see if deadalnix has some input here.
> 
> SDC does it by sticking one context in the base and one context in the child.

It's definitely different. dmd inserts at most only one context over the all derived classes. The second context pointer in the child class need to be error in SDC.

--

https://issues.dlang.org/show_bug.cgi?id=14363

--- Comment #8 from deadalnix <deadalnix@gmail.com> ---
(In reply to Kenji Hara from comment #7)
> (In reply to deadalnix from comment #6)
> > (In reply to Shammah Chancellor from comment #5)
> > > I'm not sure about that part of SDC.  I will investigate and see if deadalnix has some input here.
> > 
> > SDC does it by sticking one context in the base and one context in the child.
> 
> It's definitely different. dmd inserts at most only one context over the all derived classes. The second context pointer in the child class need to be error in SDC.

I see no reason for this limitation.

--