memeq

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

bool memeq(const uint8_t* pu1, const uint8_t* pu2, size_t n) {
    if (n >= sizeof(__m512i)) {
        size_t off;
        for (off = 0; off <= n - sizeof(__m512i); off += sizeof(__m512i)) {
            auto ymm1 = _mm512_loadu_si512((const __m512i*)(pu1 + off));
            auto ymm2 = _mm512_loadu_si512((const __m512i*)(pu2 + off));
            auto res = _mm512_cmpeq_epi8_mask(ymm1, ymm2);
            if (_kortestz_mask64_u8(res, res)) return false;
        }
        if (off < n) {
            auto ymm1 = _mm512_loadu_si512((const __m512i*)(pu1 + n - sizeof(__m512i)));
            auto ymm2 = _mm512_loadu_si512((const __m512i*)(pu2 + n - sizeof(__m512i)));
            auto res = _mm512_cmpeq_epi8_mask(ymm1, ymm2);
            if (_kortestz_mask64_u8(res, res)) return false;
        }
    }
    else if (n >= sizeof(__m256i)) {
        size_t off;
        for (off = 0; off <= n - sizeof(__m256i); off += sizeof(__m256i)) {
            auto ymm1 = _mm256_loadu_si256((const __m256i*)(pu1 + off));
            auto ymm2 = _mm256_loadu_si256((const __m256i*)(pu2 + off));
            auto res = _mm256_cmpeq_epi8(ymm1, ymm2);
            if (!_mm256_testc_si256((res), _mm256_set1_epi8(~0))) return false;
        }
        if (off < n) {
            auto ymm1 = _mm256_loadu_si256((const __m256i*)(pu1 + n - sizeof(__m256i)));
            auto ymm2 = _mm256_loadu_si256((const __m256i*)(pu2 + n - sizeof(__m256i)));
            auto res = _mm256_cmpeq_epi8(ymm1, ymm2);
            if (!_mm256_testc_si256((res), _mm256_set1_epi8(~0))) return false;
        }
    }
    else if (n >= sizeof(__m128i)) {
        {
            auto xmm1 = _mm_loadu_si128((const __m128i*)pu1);
            auto xmm2 = _mm_loadu_si128((const __m128i*)pu2);
            auto res = _mm_cmpeq_epi8(xmm1, xmm2);
            if (!_mm_test_all_ones(res)) return false;
        }
        if (n > sizeof(__m128i)) {
            auto xmm1 = _mm_loadu_si128((const __m128i*)(pu1 + n - sizeof(__m128i)));
            auto xmm2 = _mm_loadu_si128((const __m128i*)(pu2 + n - sizeof(__m128i)));
            auto res = _mm_cmpeq_epi8(xmm1, xmm2);
            if (!_mm_test_all_ones(res)) return false;
        }
    }
    else if (n >= sizeof(uint64_t)) {
        {
            auto v1 = *(const uint64_t*)pu1;
            auto v2 = *(const uint64_t*)pu2;
            if (v1 != v2) return false;
        }
        if (n > sizeof(uint64_t)) {
            auto v1 = *(const uint64_t*)(pu1 + n - sizeof(uint64_t));
            auto v2 = *(const uint64_t*)(pu2 + n - sizeof(uint64_t));
            if (v1 != v2) return false;
        }
    }
    else if (n >= sizeof(uint32_t)) {
        {
            auto v1 = *(const uint32_t*)pu1;
            auto v2 = *(const uint32_t*)pu2;
            if (v1 != v2) return false;
        }
        if (n > sizeof(uint32_t)) {
            auto v1 = *(const uint32_t*)(pu1 + n - sizeof(uint32_t));
            auto v2 = *(const uint32_t*)(pu2 + n - sizeof(uint32_t));
            if (v1 != v2) return false;
        }
    }
    else if (n >= sizeof(uint16_t)) {
        {
            auto v1 = *(const uint16_t*)pu1;
            auto v2 = *(const uint16_t*)pu2;
            if (v1 != v2) return false;
        }
        if (n > sizeof(uint16_t)) {
            auto v1 = *(pu1 + 2);
            auto v2 = *(pu2 + 2);
            if (v1 != v2) return false;
        }
    }
    else {
        if (*pu1 != *pu2) return false;
    }

    return true;
}