Thread overview
advise about safety of using move in an opAssign with __traits(isRef
Aug 16, 2019
aliak
Aug 16, 2019
Paul Backus
Aug 17, 2019
aliak
Aug 17, 2019
Paul Backus
Aug 18, 2019
aliak
August 16, 2019
Hi, I'm trying to fix a use-case where you have a wrapper template type (it's an optional) and the wrapping type has @disable this(this). And having this scenario work:

struct S {
  @disable this(this);
}
Optional!S fun() {...}

Optional!S a;
a = fun.move;

Now that won't work because of the disabled copy, but you can either:

1) Add an explicit move function in the Optional type:
struct Optional {
  auto move(ref Optional!T target) {
    import std.algorithm: move;
    return move(this, target);
  }
}

Then you  could do:
fun.move(field);

First of all I find the readability confusing, so there's no way to know if fun is moving in to field or field is moving in to fun. But OTOH it matches the phobos move(T src, T target). So conventionally maybe ok. Also it's explicit, so it makes it clear at the call site that a move is happening.

2) use isRef inside opAssign like this:

void opAssign(auto ref Optional!T lhs) {
  static if (__traits(isRef, lhs)) {
    this._value = lhs._value;
  } else {
    import std.algorithm: move;
    move(lhs._value, this._value);
  }
}

Then you could just do:

field = fun

And it'll just work since fun returns an rvalue. The second solution it seems like it'll just work correctly with the static if guard. But I'm not sure if there're any gotchas I should be aware of?

Any advice?

August 16, 2019
On Friday, 16 August 2019 at 08:07:28 UTC, aliak wrote:
> Hi, I'm trying to fix a use-case where you have a wrapper template type (it's an optional) and the wrapping type has @disable this(this). And having this scenario work:
>
> [...]
>
> 2) use isRef inside opAssign like this:
>
> void opAssign(auto ref Optional!T lhs) {
>   static if (__traits(isRef, lhs)) {
>     this._value = lhs._value;
>   } else {
>     import std.algorithm: move;
>     move(lhs._value, this._value);
>   }
> }
>
> Then you could just do:
>
> field = fun
>
> And it'll just work since fun returns an rvalue. The second solution it seems like it'll just work correctly with the static if guard. But I'm not sure if there're any gotchas I should be aware of?
>
> Any advice?

Even simpler:

void opAssign(auto ref Optional!T rhs) {
    import core.lifetime: forward;
    this._value = forward!rhs;
}
August 17, 2019
On Friday, 16 August 2019 at 14:29:26 UTC, Paul Backus wrote:
> On Friday, 16 August 2019 at 08:07:28 UTC, aliak wrote:
>> [...]
>
> Even simpler:
>
> void opAssign(auto ref Optional!T rhs) {
>     import core.lifetime: forward;
>     this._value = forward!rhs;
> }

That doesn't work with private members unfortunately :( - i.e. this.value = forward!(rhs._value) - member inaccessible.

Is there a known workaround for that? If not then should I just copy the implementation in place to be correct?
August 17, 2019
On Friday, 16 August 2019 at 08:07:28 UTC, aliak wrote:
> Hi, I'm trying to fix a use-case where you have a wrapper template type (it's an optional) and the wrapping type has @disable this(this). And having this scenario work:
>
> struct S {
>   @disable this(this);
> }
> Optional!S fun() {...}
>
> Optional!S a;
> a = fun.move;
>
> Now that won't work because of the disabled copy

Are you sure? I tried to reproduce the error you're describing, and I couldn't do it. The following compiles and runs without any issues:

struct Wrapper(T) { T value; }
struct NoCopy { @disable this(this); }

Wrapper!NoCopy fun()
{
    return Wrapper!NoCopy();
}

void main()
{
    Wrapper!NoCopy a;
    a = fun(); // already an rvalue, so moved implicitly
}

Can you give a more complete example of the problem you're having?
August 18, 2019
On Saturday, 17 August 2019 at 16:43:51 UTC, Paul Backus wrote:
> On Friday, 16 August 2019 at 08:07:28 UTC, aliak wrote:
>> Hi, I'm trying to fix a use-case where you have a wrapper template type (it's an optional) and the wrapping type has @disable this(this). And having this scenario work:
>>
>> struct S {
>>   @disable this(this);
>> }
>> Optional!S fun() {...}
>>
>> Optional!S a;
>> a = fun.move;
>>
>> Now that won't work because of the disabled copy
>
> Are you sure? I tried to reproduce the error you're describing, and I couldn't do it. The following compiles and runs without any issues:
>
> struct Wrapper(T) { T value; }
> struct NoCopy { @disable this(this); }
>
> Wrapper!NoCopy fun()
> {
>     return Wrapper!NoCopy();
> }
>
> void main()
> {
>     Wrapper!NoCopy a;
>     a = fun(); // already an rvalue, so moved implicitly
> }
>
> Can you give a more complete example of the problem you're having?

You'll get the error if you define an opAssign:

struct Wrapper(T) {
    T value;
    void opAssign(U : T)(Wrapper!U other) {
        this.value = other.value;
    }
}

struct S {
  @disable this(this);
}

Wrapper!S fun() { return Wrapper!S(S()); }

void main() {
	Wrapper!S a;
	a = fun;
}