structure binding transformation is not accurate

[viboes]

The structure binding transformation is not accurate, as the symbols are not real variables, thy are just names.

See the example

https://cppinsights.io/lnk?code=c3RydWN0IFBvaW50CnsKICBpbnQgeDsKICBpbnQgeTsKfTsKClBvaW50IHB0ezEsMn07Ci8vIEhlcmUgd2UgZ2V0IGFuIGFkZGl0aW9uYWwgb2JqZWN0IGluamVjdGVkIHRvIHdoaWNoIGF4IGFuZCBheSByZWZlci4KYXV0byBbYXgsIGF5XSA9IHB0OwoKLy8gSW4gY2FzZSBvZiBhbiByZWZlcmVuY2UgdGhlIGluamVjdGVkIG9iamVjdCBpcyBqdXN0IGEgcmVmZXJlbmNlIHRvCi8vIHRoZSBvcmlnaW5hbCBvbmUuCmF1dG8mIFthMngsIGEyeV0gPSBwdDs=&std=cpp17&rev=1.0

and this example with bitfields

https://cppinsights.io/s/a30c616b

[viboes]

The result of

https://cppinsights.io/s/d9b470ff

could be

struct Point
{
  int x;
  int y;
  // inline Point() = default;
  // inline constexpr Point(const Point &) noexcept = default;
  // inline constexpr Point(Point &&) = default;
};

void f(int, int);

Point pt = {1, 2};

void f()
{
  Point __pt12 = Point(pt);
  #define ax = __pt12.x;
  #define ay = __pt12.y;
  f(ax, ay);
  #undef ax 
  #undef ay 
}

The use of define is stillnot accurate but IMO it conveys better the intent.
Alternatively the transformation could do the name replacement :(

//...
void f()
{
  Point __pt12 = Point(pt);
  f(__pt12.x, __pt12.y);
}

Alternatively the transformation could add a comment on the variable declaration.

  // This is a workaround to have new names. You cannot take the address of these variables
  int& ax = __pt12.x; 
  int& ay = __pt12.y;

[andreasfertig]

Hello @viboes,

thank you for bringing up that issue. I agree that it is not 100% correct due to introducing new names.

While I agree, that the #define solution would be probably the most correct one, I really really hesitate implementing it. It looks not ok. Plus introducing the #undef is not trivial.

I can look into the name replacement. I sadly had to do similar things for coroutines I'm working on, but it is not trivial and usually I like to leave as much AST parts untouched as I can.

The comment is currently the easiest way to go I think. I was thinking about a annotation mode a user can turn on or off. C++ Insights for example already adds annotations for NRVO variables and other things (implicit special member functions you also have seen).

Andreas

[viboes]

I agree that the macro is not welcome.

I understand that it is not easy to do the name replacement.

If the comment is possible and easy, I believe all of us could live with.

[andreasfertig]

Hello @viboes,

I just fixed another issue related to structured binding (#425). I read the wording a couple of times without seeing wording that says that the bindings are something special. I also looked at the code in Clang and couldn't find any special treatment.

Unless somebody shows me a different interpretation of the standard or code in Clang that treats BindingDecls special, I assume that the code shown is correct and close this issue.

Andreas