fmaximum/minimum_num incorrect result for -0/+0

[gonzalobg]

See https://clang.godbolt.org/z/3W6f9faaK
Reproduced below.
When the inputs of C23 maximum/minimum_num are -0 and +0, it always returns +0.
However, clang returns -0.

#include <cmath>
#include <print>

int main() {
    using T = double;
    T nz = T(-0.);
    T pz = T(+0.);
    T r = fmaximum_num(pz,nz);
    int v = std::memcmp(&r, &pz, sizeof(double)); // if result bitwise identical to +0, then returns 0
    if (v == 0) return 0;
    std::println(stderr, "ERROR: {} != {}", r, pz); // prints 0 != -0
    return 1;
}

cc @lntue

[frederick-vs-ja]

Looks like a bug of the llvm.maximumnum.f64 intrinsic.

[el-ev]

It seems maximum_num should return +0, while minimum_num should return -0.

From the standard (N3220 Section 7.12.12), we find:

The fmaximum functions determine the maximum value of their arguments. For these functions, +0 is considered greater than −0. These functions differ from the fmaximum_num functions only in their treatment of NaN arguments.

The fminimum functions determine the minimum value of their arguments. For these functions, −0 is considered less than +0.

Reference:
https://www.open-std.org/jtc1/sc22/wg14/www/docs/n3220.pdf