I didn't see the difference of above 2 ranches. Did I miss something?

Also why not directly use memcpy?

[Cao, Minggui] I improve it here for that: it could use to move /
copy chars in the same buffer (could be partly overlap) and it could
start from

right or left.

I have checked the HV's memcpy can't fill this requirement.

1: This is not char copy, but memory copy.
2: The reason memcpy today doesn’t support this feature is because it
only work with DF=0 (REP MOVSB). We can easily extend to support DF=1.

[Cao, Minggui] sorry, I mixed the copy and move; it shall be like memmove.
Just checked the memory.c code in HV.

Is it OK as follows?
memcpy_erms --> memmove_erms (void *d, const void *s, size_t slen, bool
df_flag) as a public API?

The caller side clearly know the copy to be from head to tail address or from tail to head address.
How about to define another API like memcpy_tail_to_head or memcpy_rev ?

The caller side can choose normal memcpy or tail_to_head copy API.