Function sz_find_haswell

include/stringzilla/find.h:1063–1097

}

SZ_PUBLIC sz_cptr_t sz_find_haswell(sz_cptr_t h, sz_size_t h_length, sz_cptr_t n, sz_size_t n_length) {

    // This almost never fires, but it's better to be safe than sorry.
    if (h_length < n_length || !n_length) return SZ_NULL_CHAR;
    if (n_length == 1) return sz_find_byte_haswell(h, h_length, n);

    // Pick the parts of the needle that are worth comparing.
    sz_size_t offset_first, offset_mid, offset_last;
    sz_locate_needle_anomalies_(n, n_length, &offset_first, &offset_mid, &offset_last);

    // Broadcast those characters into YMM registers.
    sz_u32_vec_t matches_vec;
    sz_u256_vec_t h_first_vec, h_mid_vec, h_last_vec, n_first_vec, n_mid_vec, n_last_vec;
    n_first_vec.ymm = _mm256_set1_epi8(n[offset_first]);
    n_mid_vec.ymm = _mm256_set1_epi8(n[offset_mid]);
    n_last_vec.ymm = _mm256_set1_epi8(n[offset_last]);

    // Scan through the string.
    for (; h_length >= n_length + 32; h += 32, h_length -= 32) {
        h_first_vec.ymm = _mm256_lddqu_si256((__m256i const *)(h + offset_first));
        h_mid_vec.ymm = _mm256_lddqu_si256((__m256i const *)(h + offset_mid));
        h_last_vec.ymm = _mm256_lddqu_si256((__m256i const *)(h + offset_last));
        matches_vec.i32 = //
            _mm256_movemask_epi8(_mm256_cmpeq_epi8(h_first_vec.ymm, n_first_vec.ymm)) &
            _mm256_movemask_epi8(_mm256_cmpeq_epi8(h_mid_vec.ymm, n_mid_vec.ymm)) &
            _mm256_movemask_epi8(_mm256_cmpeq_epi8(h_last_vec.ymm, n_last_vec.ymm));
        while (matches_vec.u32) {
            int potential_offset = sz_u32_ctz(matches_vec.u32);
            if (sz_equal_haswell(h + potential_offset, n, n_length)) return h + potential_offset;
            matches_vec.u32 &= matches_vec.u32 - 1;
        }
    }

    return sz_find_serial(h, h_length, n, n_length);
}

SZ_PUBLIC sz_cptr_t sz_rfind_haswell(sz_cptr_t h, sz_size_t h_length, sz_cptr_t n, sz_size_t n_length) {