check_ip_header_sum.c

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

/**
  Here is my Standard-compliant, SIMD-capable version of
  https://pzemtsov.github.io/2016/11/06/bug-story-alignment-on-x86.html.
  I wonder which one is faster in a benchmark?
*/
_Bool check_ip_header_sum(const char *p, size_t size)
{
    uintmax_t sum_l = 0, sum_h = 0;
    for (size_t i = 0; i < size;)
    {
        // Endianness doesn’t really matter because the carry is circular.
        sum_l += p[i++];
        sum_h += p[i++];
    }
    sum_l += sum_h << 8;
    
    do
        sum_l = (sum_l & 0xFFFF) + (sum_l >> 16);
    while (sum_l > 0xFFFF);

    return sum_l == 0xFFFF;
}

Compiled on https://godbolt.org with x86-64 gcc (trunk) and -O3

SIMD is available for “simple” loops and produces fewer instructions with incremental iterations than decremental ones.