"ubyte[size] store = void" in std.variant

[Graham Fawcett]

On Tue, 15 Jun 2010 11:24:48 -0400, Steven Schveighoffer wrote:

> On Tue, 15 Jun 2010 11:18:07 -0400, Graham Fawcett <fawcett at uwindsor.ca>
> wrote:
> 
>> On Tue, 15 Jun 2010 10:29:56 -0400, Steven Schveighoffer wrote:
>>
>>> On Tue, 15 Jun 2010 09:04:23 -0400, Graham Fawcett
>>> <fawcett at uwindsor.ca> wrote:
>>>
>>>> Hi folks,
>>>>
>>>> The following statement appears in std.variant:
>>>>
>>>> 190    union
>>>> 191    {
>>>> 192        ubyte[size] store = void;
>>>> 193        // conservatively mark the region as pointers 194 static
>>>> if (size >= (void*).sizeof) 195            void* p[size /
>>>> (void*).sizeof]; 196    }
>>>>
>>>> The '= void' on line 192 sometimes leads to 'Error: void initializer
>>>> has no value' errors in application code. For example, this fails to
>>>> compile on DMD 2.047:
>>>>
>>>>   foreach (int v; map! "a.get!int" (variantArray(1,2,3)))
>>>>     writeln(v);
>>>>
>>>> Changing line 192 to 'ubyte[size] store;' resolves the issue.
>>>>
>>>> My question is: what is the point of the '= void' initializer here?
>>>> Would std.variant be broken if '= void' were removed?
>>>
>>> = void means don't initialize the data.  Otherwise, the
>>> compiler/runtime will fill in the data will all 0s.  However, I'm not
>>> sure how that works with a union, since you may have conflicting
>>> requirements for initialization.
>>
>> I'm not sure if it's in the spec, but a quick test results in a
>> compiler error if I declare a union with overlapping initializers.
>>
>> Simplifying the 'std.variant' case, I get the same 'void initializer
>> has no value' error like this:
>>
>>     struct foo {
>>       ubyte[] store = void;
>>     }
>>     foo z = foo();
>>
>> Is this a compiler bug?
> 
> Note, you can only use = void on a value type, not a dynamic array.  The
> variant union member is a static array, not a dynamic one.

Oops, thanks for catching that. I was trying various combinations of
types and array-sizes to look for a pattern. For the record,

  struct foo {
    enum N = 10; // or whatever
    ubyte[N] store = void;
  }
  foo z = foo();

...returns the same error.

Graham

> But actually, now that I think about it, =void is generally used in a
> function, not in a type definition.  I'm not sure = void should be
> allowed in that context.  So maybe the compiler is right to complain...
> 
> -Steve