Correct floor and ceil from truncation!

[p0nce]

Do those actually work and are correct? Found in stb_image_resize2.h
if true, could be faster than changing the rounding mode

[p0nce]

• exhaustive testing in float32

[p0nce]

• Finite number from -2e9 to 2e9 => OK, above that rounding makes no sense for a float
• Denormals => OK
• Preserve NaN-ness (warning: libc mandates to return the same NaN), test
• -0 and +0, should preserve sign, test

If x is integral, +0, -0, NaN, or infinite, x itself is returned.

[p0nce]

Complete description here:

// Spec difference with ceilf: 
// - May return non-sensical for -inf and +inf (not seen on x86_64 though)
// - Doesn't preserve the sign of -0.0
// - Doesn't preserve NaN payload, does preserve being NaN
// Works similarly to ceilf on the [-2e9, 2e9] range, including denormals.
float testceilf(float x)  // martins ceilf
{
    assert(-2e9 <= x && x <= 2e9);
    __m128 f = _mm_set_ss(x);
    __m128 t = _mm_cvtepi32_ps(_mm_cvttps_epi32(f));
    __m128 r = _mm_add_ss(t, _mm_and_ps(_mm_cmplt_ss(t, f), _mm_set_ss(1.0f)));
    return _mm_cvtss_f32(r);
}

// Spec difference with floorf: 
// - Return non-sensical values for -inf and +inf
// - Doesn't preserve the sign of -0.0
// - Doesn't preserve NaN payload, does preserve being NaN
// Works similarly to floorf on the [-2e9, 2e9] range, including denormals.
float testfloorf(float x)  // martins floorf
{
    assert(-2e9 <= x && x <= 2e9);
    __m128 f = _mm_set_ss(x);
    __m128 t = _mm_cvtepi32_ps(_mm_cvttps_epi32(f));
    __m128 r = _mm_add_ss(t, _mm_and_ps(_mm_cmplt_ss(f, t), _mm_set_ss(-1.0f)));
    return _mm_cvtss_f32(r);
}

[p0nce]

• I wonder how it fares on arm64 => people report working, so it could be used to emulate _mm_floor_ps and friends