Re: [RFC PATCH v1 0/5] nolibc x86-64 string functions

[Willy Tarreau]

On Fri, Sep 01, 2023 at 11:34:18AM +0000, David Laight wrote:
> From: Ammar Faizi <ammarfaizi2@xxxxxxxxxxx>
> > Sent: 30 August 2023 14:57
> > 
> > This is an RFC patchset for nolibc x86-64 string functions. There are 5
> > patches in this series.
> > 
> > ## Patch 1-3: Use `rep movsb`, `rep stosb`, and `rep cmpsb` for:
> >     - memcpy() and memmove()
> >     - memset()
> >     - memcmp()
> > respectively. They can simplify the generated ASM code.
> > 
> ...
> > After this series:
> > ```
> >   000000000000140a <memmove>:
> >     140a: 48 89 f8              mov    %rdi,%rax
> >     140d: 48 89 d1              mov    %rdx,%rcx
> >     1410: 48 8d 7c 0f ff        lea    -0x1(%rdi,%rcx,1),%rdi
> >     1415: 48 8d 74 0e ff        lea    -0x1(%rsi,%rcx,1),%rsi
> >     141a: fd                    std
> >     141b: f3 a4                 rep movsb %ds:(%rsi),%es:(%rdi)
> >     141d: fc                    cld
> >     141e: c3                    ret
> 
> Isn't that completely broken?
> 
> You need to select between forwards and backwards moves.
> Since forwards moves are preferred it is best to do
> 	if (dst - src < len)
> 		backards_copy()
> 	else
> 		formwards_copy()
> 
> 	David

You're completely right indeed, reminds me about the copy_up/copy_down
that were not used anymore :-)

Willy