std::tuple<> is not trivially constructible

[llvmbot]

Bugzilla Link	41714
Resolution	FIXED
Resolved on	Jun 11, 2019 08:06
Version	unspecified
OS	Linux
Reporter	LLVM Bugzilla Contributor
CC	@ldionne,@mclow
Fixed by commit(s)	r363075

Extended Description

Clang -std=c++17 compiles the following if using libstdc++:

#include
static_assert(std::is_trivially_constructible_v<std::tuple<>>);

But adding -stdlib=libc++ causes

test.cc:3:1: error: static_assert failed due to requirement 'std::is_trivially_constructible_v<std::__1::tuple<>>

[llvmbot]

assigned to @ldionne

[mclow]

A few minutes of looking through the standard did not turn up any requirement that tuple<> be trivially constructible.

trivially destructible, yes.

4 If is_trivially_destructible_v is true for all Ti, then the destructor of tuple is trivial.

constexpr, yes.

The defaulted move and copy constructor of tuple<> shall be constexpr functions.

Can you tell me where you believe this requirement comes from?

[llvmbot]

I have no idea if the standard requires this, but it would be a nice to have extension if it doesn't. Specially considering that libstdc++ has it.

[mclow]

I have no idea if the standard requires this, but it would be a nice to have
extension if it doesn't. Specially considering that libstdc++ has it.

Generally speaking, I'm philosophically opposed to "things that everyone does, but aren't in the standard".

If this is a good thing, then we should write it into the standard.

[ldionne]

I think it's a reasonable thing to do as a matter of QOI. The change is very simple, too, and it's the natural way to write that default constructor IMO.

Review here: https://reviews.llvm.org/D62618

[ldionne]

Fixed in r363075.