[Moved to DR at the February, 2014 meeting.]

Proposed resolution (September, 2013):

Change 7.3.2 [conv.lval] paragraph 2 as follows:

When an lvalue-to-rvalue conversion occurs in an unevaluated operand or a subexpression thereof ( Clause 7 [expr]) is applied to an expression e, and either

• e is not potentially evaluated, or

• the evaluation of e results in the evaluation of a member ex of the set of potential results of e, and ex names a variable x that is not odr-used by ex (6.3 [basic.def.odr]),

the value contained in the referenced object is not accessed. [Example:

  struct S { int n; };
  auto f() {
    S x { 1 };
    constexpr S y { 2 };
    return [&](bool b) { return (b ? y : x).n; };
  }
  auto g = f();
  int m = g(false); // undefined behavior due to access of x.n outside its lifetime
  int n = g(true);  // OK, does not access y.n

—end example] In all other cases, the result of the conversion...

Consider an example like the following:

  struct Base {
    virtual int call() = 0;
  };
  Base *foo() {
    constexpr int x = 0;
    struct Local : Base {
      virtual int call() { return x; }
    };
    static Local local;
    return &local;
  }

  int main() {
    return foo()->call();
  }

While the likely intention is that the lvalue-to-rvalue conversion of the block-scope constant is implemented by using the value of the constant expression in place of reading from storage, it seems that the wording of 7.3.2 [conv.lval] paragraph 2 does not prevent this program from being subject to undefined behaviour caused by lifetime violation. In particular, it seems that a name expression that appears in a potentially-evaluated expression such that the object named is not odr-used (by that instance of the name) may still be evaluated, in theory, as an lvalue through which the object named or a subobject thereof is accessed.