Lambda in branch not taken of constant expression: Who is right?

Question

I tried to compile the following C++11 code with mixed results.

struct NoTemplate {
    static constexpr auto (*foo)() = false ? +[]{} : nullptr;
};

NoTemplate no_inst;


template<typename>
struct YesTemplate {
    static constexpr auto (*foo)() = false ? +[]{} : nullptr;
};

YesTemplate<float> yes_inst;

clang: Compiles NoTemplate sucessfully; gives error: a lambda expression may not appear inside of a constant expression on YesTemplate.
gcc: Compiles both successfully
msvc: Crashes.
icc: Crashes (we have a winner!)

What is the correct result? I see some standard language suggesting non-constant expressions should be OK in the false branch of short-circuiting operators in constant expressions, but IANALL.

By crashes do you mean the compiler crashes or the program crashes? — user4581301, Dec 19 '17 at 22:06
In the latest VC (19.12.something) with the standard set to c++17 and warnings set to level 4 I get ambiguous operator on the '+' — SoronelHaetir, Dec 19 '17 at 22:15
Looks like a clang bug, clang HEAD does not provide a diagnostic [see it ive](https://wandbox.org/permlink/2I7TE8WOTzZvmeZC) the language [supports lambdas in unevaluated portion of conditional](https://stackoverflow.com/q/22232164/1708801) — Shafik Yaghmour, Dec 19 '17 at 22:21
Well, we can definitely say compiler crashes are always a bug and never intended behavior. :) Nice job finding multiple crashes though. — GManNickG, Dec 19 '17 at 23:08
@asimes unary `operator+` on the lamdba that effectively converts it to a function pointer. — kmdreko, Dec 21 '17 at 03:44

xskxzr · Answer 1 · 2017-12-21T03:04:34.940

1

This program is well-formed.

Note after C++17, a lambda expression may be accepted in a core constant expression even if it would be evaluated. You can see the proposed paper N4487 for detail.

edited Dec 21 '17 at 03:04

answered Dec 20 '17 at 14:40

xskxzr

12,442
12
37
77

Lambda in branch not taken of constant expression: Who is right?

1 Answers1