Structured bindings seem to allow non-const references to a bitfield

Clang Frontend
1

I wasn't sure whether to bring this up here, or to bring it up on the reflector, but I'm not sure if this is a compiler issue or a standards issue. I read through the structured binding wording, and don't see anything one way or the other, and GCC7 and Clang trunk both allow the following:

struct F {
  unsigned a : 1;
  unsigned b : 2;
  unsigned c : 8;
};

int main() {
  F f;
  f.a = 0;
  f.b = 1;
  f.c = 4;

  auto &b = f.b; // correctly errors on non-const ref to a bitfield

  auto& [d,e,g] = f; // all are references to a bitfield
  d = 3; // Works, but warns about truncation
  c = 3; // Works, no diagnostic
}

Is this intended? Is this justified by the standard?
-Erich

2

I wasn't sure whether to bring this up here, or to bring it up on the
reflector, but I'm not sure if this is a compiler issue or a standards
issue. I read through the structured binding wording, and don't see
anything one way or the other, and GCC7 and Clang trunk both allow the
following:

struct F {
  unsigned a : 1;
  unsigned b : 2;
  unsigned c : 8;
};

int main() {
  F f;
  f.a = 0;
  f.b = 1;
  f.c = 4;

  auto &b = f.b; // correctly errors on non-const ref to a bitfield

  auto& [d,e,g] = f; // all are references to a bitfield

The only reference here is an unnamed reference to `f`. None of `d`, `e`,
`g` is a reference. `d` is another name for `f.a`, and similarly for `e`
and `g`.

  d = 3; // Works, but warns about truncation
  c = 3; // Works, no diagnostic
}

Is this intended? Is this justified by the standard?

The behavior you describe appears correct (and unsurprising to me), apart
from the assumption that these are references; they're "structured
bindings", not references.

-- James

3

Ok, I guess I didn’t (don’t completely?) ‘get’ that there is a difference between “structured binding alias” and “reference”. Additionally, if you simply do “auto”, you get copies, right? So those structured bindings actually introduce new variables, right?

4

Ah, I actually found a more recent version of P0217 (R2 in this case) that mentions bitfields. Thanks for the clarification!

-Erich

5

You get (at most) one copy, of the value that's being decomposed, and it is
unnamed.

-- James

6

AH! I get it now. The ‘auto’/’const auto’/’auto&’ actually applies to an unnamed Structured Binding object, not as a qualifier to the individual items. A structured binding is more akin to a named tuple rather than introducing values.

I missed that mental model, thank you very much!

-Erich