Trivially-constructible-from

Today I learned:

using u64 = long long;
using u64b = long;
static_assert(sizeof (u64) == 8);
static_assert(sizeof (u64b) == 8);

void foo(u64 *p, u64 *q, int n)
{
    std::copy(p, p+n, q);
}

void bar(u64 *p, u64b *q, int n)
{
    std::copy(p, p+n, q);
}

Both Clang and GCC were smart enough to optimize foo into 8 or 9 instructions ending with a tail-call to memmove.

bar generates 30 instructions on GCC; 82 on Clang.

It would be cool if library vendors could just check is_trivially_constructible_v<u64, u64b> and optimize to a memmove in that case. But we can’t have nice things:

assert(is_trivially_constructible_v<u64, u64b>);
// Yay!

using u16 = short;
assert(is_trivially_constructible_v<u64, u16>);
// What the...

assert(is_trivially_constructible_v<u64, double>);
// ...oh geez.

Yep, is_trivially_constructible returns always-true for scalar types. Because the standard says

[the construction] is known to call no operation that is not trivial

where “trivial” is a concept that is never defined anywhere in the Standard (except for special member functions, and u64 doesn’t have member functions). Vendors have interpreted this as meaning that scalar types should always count as “trivially constructible,” because initializing a scalar type never “calls” any operation (although it may still perform arbitrarily complex math).

This is why we can’t have nice things.


Unrelatedly, GCC currently believes that is_trivially_constructible_v<int,void*> == true and simultaneously that is_constructible_v<int,void*> == false. This is a bug.

Posted 2018-07-03