Compile time floating point division by zero in C++

It is well known, if one divides a floating-point number on zero in run-time then the result will be either infinity or not-a-number (the latter case if the dividend was also zero). But is it allowed to divide by zero in C++ constexpr expressions (in compile time), e.g.

#include <iostream>

int main() {
    double x = 0.;
    // run-time division: all compilers print "-nan"
    std::cout << 0./x << std::endl; 

    // compile-time division, diverging results
    constexpr double y = 0.;
    std::cout << 0./y << std::endl;
}

In this program the first printed number is obtained from division in run-time, and all compilers are pretty consistent in printing -nan. (Side question: why not +nan?)

But in second case the compilers diverge. MSVC simply stops compilation with the error:

error C2124: divide or mod by zero

GCC still prints -nan, while Clang changes the sign printing “positive” nan, demo: https://gcc.godbolt.org/z/eP744er8n

Does the language standard permit all three behaviors of the compilers for compile-time division: 1) reject the program, 2) produce the same division result as in run-time, 3) produce a different (in sign bit) result?