#include <stdint.h>
#include <stddef.h>
#include <emmintrin.h>
#include <tmmintrin.h>

/// Aligned search max item
static inline uint8_t search_max_octet_aligned(const __m128i *src, size_t length) {
    __m128i vmax = _mm_set_epi32(0, 0, 0, 0);
    __m128i mj;
    while (length --> 0) {
        mj   = _mm_load_si128(src++);
        vmax = _mm_max_epu8  (vmax, mj);
    }
    vmax = _mm_max_epu8(vmax, _mm_alignr_epi8(vmax, vmax, 1));
    vmax = _mm_max_epu8(vmax, _mm_alignr_epi8(vmax, vmax, 2));
    vmax = _mm_max_epu8(vmax, _mm_alignr_epi8(vmax, vmax, 4));
    vmax = _mm_max_epu8(vmax, _mm_alignr_epi8(vmax, vmax, 8));
    return _mm_extract_epi16(vmax, 0) & 0xFF;
}

uint8_t search_max_octet(const uint8_t *src, size_t length) {
    uint8_t result = 0;
    while (length && (((uintptr_t)src) & 0xF)) {
       result = result < src[0] ? src[0] : result;
       src++;
       length--;
    }
    if (length > 16) {
       size_t rl = length & ~0xF;
       result = search_max_octet_aligned((const __m128i *)src, rl >> 16);
       src += rl;
       length -= rl;
    }
    for (unsigned i = 0; i < length; i++)
       result = result < src[i] ? src[i] : result;
    return result;
}