Slicing XOR hashing

slice_xor.c

#include <stdint.h>
#include <stdlib.h>
#include <string.h>

#include <x86intrin.h>
#include <avx2intrin.h>

#ifdef __cplusplus
extern "C" {
void slice_xor(const unsigned char *data, size_t len, uint64_t *p1, uint64_t *p2);
}
#endif

void slice_xor(const unsigned char *data, size_t len, uint64_t *p1, uint64_t *p2) {
        __m256i res = { 0 };

        while(len >= 32) {
                __m256i current = _mm256_stream_load_si256((const __m256i *)data);
                res = _mm256_xor_si256(res, current);
                data += 32;
                len -= 32;
        }

        if(len > 0) {
                __m256i tmp = { 0 };
                memcpy(&tmp, data, len);
                res = _mm256_xor_si256(res, tmp);
        }

        __m128i a = _mm256_extracti128_si256(res, 0),
                b = _mm256_extracti128_si256(res, 1);

        union { __m128i v; uint64_t i64[2]; } res2 = { .v = _mm_xor_si128(a, b) };
        *p1 = res2.i64[0];
        *p2 = res2.i64[1];
}