rsmith added a comment. In https://reviews.llvm.org/D46845#1099732, @mclow.lists wrote:
> >> One way we could deal with this is by adding an attribute to the compiler > >> to indicate "the const is a lie", that we can apply to std::pair::first, > >> with the semantics being that a top-level const is ignored when > >> determining the "real" type of the member (and so mutating the member > >> after a const_cast has defined behavior). This would apply to all > >> std::pairs, not just the node_handle case, but in practice we don't > >> optimize on the basis of a member being declared const *anyway*, so this > >> isn't such a big deal right now. > > > > I'm a fan of this idea, this would also have the bonus of fixing some > > pre-existing type-punning between pair<const K, V> and pair<K, V> (see > > __hash_value_type and __value_type in <unordered_map> and <map>). > > This was //supposed// to be taken care of by [class.mem]/21 (common initial > sequence), but that foundered on "standard-layout". Also, the common initial sequence rule only allows loads through the wrong type, not stores, and in any case does not alter the rule that modifying a const object results in UB. https://reviews.llvm.org/D46845 _______________________________________________ cfe-commits mailing list cfe-commits@lists.llvm.org http://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
