It's not possible to store bit masks in SIMD register for logical operations.

[juj]

Build the following code:

#include <xmmintrin.h>
#include <stdio.h>
#include <string>

std::string tostr(__m128 m)
{
    unsigned int u[4];
    _mm_store_ps((float*)u, m);
    char str[256];
    sprintf(str, "[%X,%X,%X,%X]", u[3], u[2], u[1], u[0]);
    return str;
}

int main()
{
    unsigned int u[4] = { 0xFFFFFFFFU, 0xFFFFFFFFU, 0xFFFFFFFFU, 0xFFFFFFFFU };
    __m128 m1 = _mm_load_ps((float*)&u);
    printf("%s\n", tostr(m1).c_str());
}

Running in FF Nightly prints out [7FC00000,7FC00000,7FC00000,7FC00000], whereas it should print [FFFFFFFF,FFFFFFFF,FFFFFFFF,FFFFFFFF].

[sunfishcode]

Yes, this does look like NaN caonicalization in the polyfill, as you noted in tc39/ecmascript_simd#38 (comment)

The only ways I can think of to fix this in the polyfill involve making float32x4 hold raw bit patterns and convert them to actual float32 values on demand, but that would add a bunch of complexity and slow it down.

Native implementations can avoid this problem because a plain SIMD load won't involve loading scalar values, so it won't require NaN canonicalization.

[juj]

Added a test for this in the above commit. This passes in node.js which uses the polyfill, but does not pass in SpiderMonkey, which should not be using the polyfill, so I don't think this is a polyfill issue, but related to SpiderMonkey instead?

[juj]

Like said earlier, this is not a polyfill issue. LLVM is failing here.

Building with em++ s.cpp -o s.html -msse -O0, or -O2 or -O3 and running in current Nightly works, but building with O1 does not, and the NaNs get canonicalized. The problem is that LLVM is generating optimizations that break under canonicalization.

[juj]

Fixed in the above PR, works in current incoming.