gist:dabd69b249a102c2c57c

## gistfile1.cpp
// Essentially the same algorithm as
//   https://github.com/ispc/ispc/blob/master/builtins/target-sse2-common.ll#L153
// but slightly better impl (orps on magic const instead of two xorps).
//
// NOTE: this really is not tested, and it's likely to lose a bit of precision
// for x in [2^22,2^23).
//
// NOTE 2: Yup, confirmed, it doesn't do the right thing for large integer values
// representable as float.
__m128 round_to_int(__m128 x)
{
    __m128 sign = _mm_and_ps(x, _mm_set1_ps(-0.0f));
    __m128 magic = _mm_or_ps(sign, _mm_set1_ps(8388608.0f)); // 2^23
    __m128 round1 = _mm_add_ps(x, magic);
    __m128 round2 = _mm_sub_ps(round1, magic);
    return round2;
}

__m128 floor(__m128 x)
{
    __m128 y = round_to_int(x);
    __m128 cmp = _mm_cmplt_ps(x, y); // x < round(x)?
    __m128 fudge = _mm_and_ps(cmp, _mm_set1_ps(1.0f)); // 1.0f if x < round(x)
    __m128 result = _mm_sub_ps(y, fudge);
}

// This variant of "round" should work, though.
// But I think it's more work than going through ints.
// (Still completely untested, mind.)
__m128 round_to_int_fixed(__m128 x)
{
    __m128 magic = _mm_set1_ps(8388608.0f); // 2^23
    __m128 xabs = _mm_and_ps(x, _mm_castsi128_ps(_mm_set1_epi32(0x7fffffff)));
    __m128 sign = _mm_xor_ps(x, xabs);
    __m128 smallenough = _mm_cmplt_ps(xabs, magic);
    __m128 bias = _mm_and_ps(smallenough, magic);
    __m128 round1 = _mm_add_ps(xabs, bias);
    __m128 round2 = _mm_sub_ps(round1, bias);
    __m128 round3 = _mm_xor_ps(round2, sign);
    return round3;
}