Function sz_copy_sve

include/stringzilla/memory.h:1249–1283

}

SZ_PUBLIC void sz_copy_sve(sz_ptr_t target, sz_cptr_t source, sz_size_t length) {
    sz_size_t vec_len = svcntb(); // Vector length in bytes

    // When the buffer is small, there isn't much to innovate.
    if (length <= vec_len) {
        // Small buffer case: use mask to handle small writes
        svbool_t mask = svwhilelt_b8((sz_u64_t)0ull, (sz_u64_t)length);
        svuint8_t data = svld1_u8(mask, (sz_u8_t *)source);
        svst1_u8(mask, (sz_u8_t *)target, data);
    }
    // Slightly larger buffers
    else if (2 * length <= vec_len) {
        svbool_t mask_first = svptrue_b8();
        svbool_t mask_second = svwhilelt_b8((sz_u64_t)vec_len, (sz_u64_t)length);
        svuint8_t data_first = svld1_u8(mask_first, (sz_u8_t *)(source));
        svuint8_t data_second = svld1_u8(mask_second, (sz_u8_t *)(source + vec_len));
        svst1_u8(mask_first, (sz_u8_t *)(target), data_first);
        svst1_u8(mask_second, (sz_u8_t *)(target + vec_len), data_second);
    }
    // For medium-sized buffers, use unidirectional traversal without non-temporal operations
    else {
        // Main loop: full vector copies
        for (; length >= vec_len; source += vec_len, target += vec_len, length -= vec_len) {
            svuint8_t data = svld1_u8(svptrue_b8(), (sz_u8_t const *)source);
            svst1_u8(svptrue_b8(), (sz_u8_t *)target, data);
        }

        // Tail: single masked copy for remainder
        if (length) {
            svbool_t mask = svwhilelt_b8((sz_u64_t)0ull, (sz_u64_t)length);
            svuint8_t data = svld1_u8(mask, (sz_u8_t const *)source);
            svst1_u8(mask, (sz_u8_t *)target, data);
        }
    }
}

SZ_PUBLIC void sz_move_sve(sz_ptr_t target, sz_cptr_t source, sz_size_t length) {
    sz_size_t vec_len = svcntb(); // Vector length in bytes