Possible incorrect code generation when using variable templates

[jonas-eschmann]

Hello,

I'm encountering an issue with member variables not being initialised properly after moving from GCC to Clang. I managed to boil it down to the following minimal example:

template <typename T>
struct A{
    int x = 10;
};

template <typename T>
struct B{
    A<T> a;
};

template <typename T>
A<T> a;

template <typename T>
auto aa = a<T>;

template <typename T>
B<T> b = {
    aa<T>
};

int main(){
    return b<float>.a.x;
}

Same example on the compiler explorer

I'm expecting it to return 10 (which GCC does) but Clang returns 0

Best regards!
Jonas

[llvmbot]

@llvm/issue-subscribers-clang-frontend

[shafik]

@erichkeane wdyt? There is a lot going on, MSVC seems to agree with clang but I am not sure if I am missing something subtle here.

[royjacobson]

Looks wrong to me as well.

[erichkeane]

I have no idea here :) I find myself wondering if there is some sort of non-guaranteed initialization order happening here? If you look at the assembly, a<float> IS initialized to be 10, though all the otherse are zero-init'ed. Its a bit confusing why it wouldn't be 10 for the result here, but I have no idea whether it is something the standard is unclear about (particularly because MSVC agrees with us).

[royjacobson]

Ah, that explains it. If you look at the global guard variable guards I think it's pretty clear :)

BTW, if you make a<T> a constexpr var you get the expected result.

[shafik]

It looks like this changed recently and now output the expected value in trunk: https://godbolt.org/z/z3GeohMxv