Miscompilation in _mm512_reduce_add_pd, assumes values not NaN

[orlp]

The following example should obviously compile to a panic, as the documentation of _mm512_reduce_add_pd mentions nothing about NaNs not being respected:

#![feature(stdsimd)]
use core::arch::x86_64::*;

#[no_mangle]
unsafe fn test() -> f64 {
    let ret = _mm512_reduce_add_pd(_mm512_set_pd(0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, f64::NAN));
    if ret.is_nan() {
        panic!("hi")
    }
    ret
}

However, when compiled with -C opt-level=3 -C target-cpu=cannonlake on the latest nightly there is no panic generated, despite the output being NaN regardless:

.LCPI0_0:
        .quad   0x7ff8000000000000
test:
        vmovsd  xmm0, qword ptr [rip + .LCPI0_0]
        ret

[apiraino]

WG-prioritization assigning priority (Zulip discussion).

@rustbot label -I-prioritize +P-medium +T-libs +requires-nightly