Jump to Table of Contents Pop Out Sidebar

P1381R1
Reference capture of structured bindings

Published Proposal,

This version:
https://wg21.link/p1381r1
Author:
Nicolas Lesser (blitzrakete[at]gmail[dot][com])
Audience:
EWG
Project:
ISO/IEC JTC1/SC22/WG21 14882: Programming Language — C++

Abstract

Analysis of reference capture of structured bindings.

1. Changes

Revision 1:

2. Introduction

In San Diego, EWG decided to allow capture by value for structured bindings. The following polls were taken:

  1. Capturing structured bindings by value either by default capture or by name capture always copies the value of the structured binding

| SF | F | N | A | SA |

| 12 | 14| 1 | 2 | 3 |

  1. Capturing structured bindings by value, either by default capture or by named capture, always copies whole/underlying object

| SF | F | N | A | SA |

| 2 | 3 | 5 | 6 | 8 |

  1. Capturing structured bindings by value by default capture captures the underlying object and by named capture only the named bindings.

| SF | F | N | A | SA |

| 0 | 3 | 3 | 8 | 14 |

The first poll got consensus, thus allowing:

struct Foo { int a; };
void f() {
    auto[a] = Foo();
    auto f1 = [a] { return a; }; // ok, captures 'a'
    auto f2 = [=] { return a; }; // ok, captures 'a'
}

However, there were concerns about capturing structured bindings by reference.

3. Problem

Bit-fields.

struct Foo { int a : 1; };
void f() {
    auto[a] = Foo();
    // 'a' is a bit-field!

    auto f1 = [&a] { return a; }; // capturing a bit-field by reference?!?
}

Since EWG agreed on the behavior of by-value capture of structured bindings, it would make sense to allow by-reference capture to behave the same, i.e. that only a reference to the bindings capture are formed and not to the underlying object.

4. What can we do?

Bit-fields have some funny behavior. There isn’t a type for them for example. One important difference is that you can’t take the address of a bit-field, so it follows that a reference to a bit-field is also ill-formed.

struct Foo { int a : 1; };
void f(int&);

int main() {
  Foo foo;
  f(foo.a); // ill-formed, reference to bit-field
}

As such, I’m proposing that reference capture of a bit-field remains ill-formed. I see no reason why this proposal should change this.

struct Foo { int a : 1; int b; };

int main() {
  auto[a, b] = Foo();

  auto f1 = [&] { return a; };      // still ill-formed with this proposal
  auto f2 = [&a = a] { return a; }; // always was ill-formed; no change
  auto f3 = [&a] { return a; };     // still ill-formed with this proposal

  auto g1 = [&] { return b; };      // previously ill-formed; well-formed with this proposal
  auto g2 = [&b = b] { return b; }; // always was well-formed; no change
  auto g3 = [&b] { return b; };     // previously ill-formed; well-formed with this proposal
}

Every other structured binding has a "valid" type (a type which we can reference); bit-fields are the only exception to this rule.

5. Proposed wording

This wording is based on P1091r3.

Change [expr.prim.lambda.capture]p12 (7.5.5.2) as follows:

A bit-field , a structured binding, or a member of an anonymous union shall not be captured by reference.