On Monday, 10 October 2016 at 10:45:37 UTC, Sönke Ludwig wrote:
> Would have been nice in theory to have real void initialization of course, plus it was there for working around that (fixed?) issue with slow compilation times for large static arrays, but there is probably no real reason now to keep it.

Here is the ER for using `= void` field initializers.
[Issue 11331 – Inefficient initialization of struct with members = void](https://issues.dlang.org/show_bug.cgi?id=11331)

Am 10.10.2016 um 12:45 schrieb Sönke Ludwig:
> Am 10.10.2016 um 12:20 schrieb Martin Nowak:
>> On Monday, 10 October 2016 at 09:03:53 UTC, Sönke Ludwig wrote:
>>> Of course, the new error is more restrictive than it should be, namely
>>> if the uninitialized pointer field gets written before the first read,
>>> it would still be safe.
>>
>> That's surprising b/c void initializers for struct fields didn't use to
>> work.
>
> Hm, thanks for the hint - if that's still the case, that leads to the
> very simple workaround of simply removing the "= void". Would have been
> nice in theory to have real void initialization of course, plus it was
> there for working around that (fixed?) issue with slow compilation times
> for large static arrays, but there is probably no real reason now to
> keep it.
>
>> I need to research the intent behind this to say sth. detailed, though
>> usually an shouldn't break working code, only deprecate it.
>

Okay, I stumbled over another occurrence of this:

Bugzilla 16195: delete should be @system
  <- This makes it impossible to use `scope` variables in @safe scopes. The whole scope/function needs to be marked trusted now.

Bugzilla 14496: void initialization of member with indirections must not be @safe
  <- This makes it impossible to use std.typecons.scoped as an alternative, because it requires void initialization due to its disabled default constructor, or it would again affect the @safety of the whole surrounding scope.

I think both need to be adjusted to use a deprecation warning instead of an error. But the first one in conjunction with `scope` variables should be legal anyway, as long as the variable doesn't leave the scope (e.g. DIP1000). Maybe it makes sense to lower this to "scope (exit) () @trusted { delete var; } ();" instead of "scope (exit) delete var;" for now (making sure that the stack allocation optimization still works).