uintptr_t LineOffsets(const char* source, uintptr_t size, uint32_t* buffer)
{
    const __m128i crvec = _mm_set1_epi8('\r');
    const __m128i nlvec = _mm_set1_epi8('\n');

    uintptr_t bufferOffset = 0;
    for (uintptr_t offset = 0; offset < size;)
    {
        __m128i srcvec = _mm_loadu_si128((const __m128i*)(source + offset));

        uint32_t crmask = _mm_movemask_epi8(_mm_cmpeq_epi8(srcvec, crvec));
        uint32_t nlmask = _mm_movemask_epi8(_mm_cmpeq_epi8(srcvec, nlvec));

        uint32_t mask = (nlmask | ((crmask << 1) & ~nlmask) >> 1) << 1;

        for (uintptr_t i = 0; i < 15; ++i)
            buffer[bufferOffset + i] = offset + _tzcnt_u32(_pdep_u32(1 << i, mask));

        bufferOffset += _popcnt32(mask);
        offset += 15;
    }

    return bufferOffset;
}