Document Number





Programming Language C++


Evolution Working Group


This paper proposes allowing dynamic_cast and polymorphic typeid in constant expressions.


P1064, accepted at Rapperswil, relaxed the restriction of virtual function calls in constant expressions. These are now allowed, a capability made possible by the fact that compilers are effectively required to track the dynamic type of constant expressions.

It would be natural, for the same underlying reasons, to also drop the restriction of dynamic_cast and typeid (applied to polymorphic glvalues) not appearing in constant expressions. P1064 mentions this in its Further Work section, and it is our impression that the EWG in Rapperswil was generally approving of the idea.

We conservatively did not propose this in P1064 (our mistake) because we did not have a motivating example, but we do now.

Proposed Changes

(All edits are relative to N4778.)

Remove the 2.15, 2.19 bullets of 7.7 [expr.const] p2:

(2.15) — a dynamic cast (;

(2.19) — a typeid expression ( whose operand is a glvalue of a polymorphic class type;

Change bullet 24 of [expr.const]p2:

(2.24) — a throw-expression (7.6.17) or a dynamic cast ( or typeid expression ( that would throw an exception;

Motivating Example

The papers P1196, P1197, P1198 all propose improvements to std::error_category that add new members (P1196) or new virtual functions (P1197, P1198) to it.

Unfortunately, changing a standard polymorphic type is an ABI break, and library implementers are increasingly unwilling to break ABI even in small quantities.

This presents a significant and frustrating obstacle to our efforts to evolve and improve the standard library.

One alternative approach is to avoid changing std::error_category, but instead derive std::error_category_v2 from it, which could now hold the additional members and virtuals.

With this done, std::error_code could use dynamic_cast to check whether it has been constructed from a "new" category, and adjust accordingly.

This, however, clashes with the proposal in P1195 to add constexpr to error_code and friends. We’re again forced to choose between one desirable improvement and another, and can’t have nice things.

The ability to use dynamic_cast in constant expressions will solve this problem.

-- end