2023-09-02 07:43:39

by Willy Tarreau

[permalink] [raw]
Subject: Re: [RFC PATCH v2 1/4] tools/nolibc: x86-64: Use `rep movsb` for `memcpy()` and `memmove()`

On Sat, Sep 02, 2023 at 01:11:06PM +0700, Ammar Faizi wrote:
> On Sat, Sep 02, 2023 at 01:07:50PM +0700, Alviro Iskandar Setiawan wrote:
> > Btw, sir, this can be simplified more by merging the forward copy
> > path, only using two "rep movsb" for both memmove() and memcpy()
> > should be enough?
> > ```
> > __asm__ (
> > ".section .text.nolibc_memmove_memcpy\n"
> > ".weak memmove\n"
> > ".weak memcpy\n"
> > "memmove:\n"
> > "movq %rdx, %rcx\n"
> > "movq %rdi, %rdx\n"
> > "movq %rdi, %rax\n"
> > "subq %rsi, %rdx\n"
> > "cmpq %rcx, %rdx\n"
> > "jnb __nolibc_forward_copy\n"
> > "leaq -1(%rdi, %rcx, 1), %rdi\n"
> > "leaq -1(%rsi, %rcx, 1), %rsi\n"
> > "std\n"
> > "rep movsb\n"
> > "cld\n"
> > "retq\n"
> >
> > "memcpy:\n"
> > "movq %rdi, %rax\n"
> > "movq %rdx, %rcx\n"
> > "__nolibc_forward_copy:\n"
> > "rep movsb\n"
> > "retq\n"
> > );
> > ```
>
> Looks good. I'll apply that change.

Note that in this case we simply don't need a special
version of memcpy(), memmove() is always OK for this,
so you can simplify this further by starting with:

memcpy:
memmove:
...

Willy