From: Ammar Faizi
> Sent: 01 September 2023 14:06
...
> > You're completely right indeed, reminds me about the copy_up/copy_down
> > that were not used anymore :-)
> 
> I'm an idiot, will fix that. Another attempt as suggested below:
> 
> __asm__ (
> ".section .text.nolibc_memmove\n"
> ".weak memmove\n"
> "memmove:\n"
> "    movq    %rdx, %rcx\n"
> "    movq    %rdi, %rdx\n"
> "    movq    %rdi, %rax\n"

You seem to have confused yourself about whether you are using %eax or %edx.

> "    subq    %rsi, %rdx\n"
> "    cmpq    %rcx, %rdx\n"
> "    jnb     .Lforward_copy\n"

I think I'd fall through to the forwards copy
and not worry about replicating the 'reps movsb' and 'ret'.
IIRC 'cld' can be slow as well.

> "    leaq    -1(%rdi, %rcx, 1), %rdi\n"
> "    leaq    -1(%rsi, %rcx, 1), %rsi\n"
> "    std\n"
> ".Lforward_copy:\n"
> "    rep movsb\n"
> "    cld\n"
> "    ret\n"
> );
> 
> --
> Ammar Faizi

	David

-
Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK
Registration No: 1397386 (Wales)

On Fri, Sep 01, 2023 at 02:23:28PM +0000, David Laight wrote:
> From: Ammar Faizi
> > Sent: 01 September 2023 14:06
...
> > __asm__ (
> > ".section .text.nolibc_memmove\n"
> > ".weak memmove\n"
> > "memmove:\n"
> > "    movq    %rdx, %rcx\n"
> > "    movq    %rdi, %rdx\n"
> > "    movq    %rdi, %rax\n"
> 
> You seem to have confused yourself about whether you are using %eax or %edx.

What do you mean? They're all 64-bit pointers.

What I know is that the %rdx will be clobbered by "subq %rsi, %rdx"
below and the %rax should be return value. That's why I copy the %rdi
twice. memmove() returns the dst pointer. Did I miss something?

> > "    subq    %rsi, %rdx\n"
> > "    cmpq    %rcx, %rdx\n"
> > "    jnb     .Lforward_copy\n"
> 
> I think I'd fall through to the forwards copy
> and not worry about replicating the 'reps movsb' and 'ret'.
> IIRC 'cld' can be slow as well.

Alright, I will avoid cld for the forward copy.

> > "    leaq    -1(%rdi, %rcx, 1), %rdi\n"
> > "    leaq    -1(%rsi, %rcx, 1), %rsi\n"
> > "    std\n"
> > ".Lforward_copy:\n"
> > "    rep movsb\n"
> > "    cld\n"
> > "    ret\n"
> > );

-- 
Ammar Faizi

From: Ammar Faizi
> Sent: 01 September 2023 15:42
...
> > > "    movq    %rdx, %rcx\n"
> > > "    movq    %rdi, %rdx\n"
> > > "    movq    %rdi, %rax\n"
> >
> > You seem to have confused yourself about whether you are using %eax or %edx.
> 
> What do you mean? They're all 64-bit pointers.

%ax, %eax, %rax - what is the difference :-)

> What I know is that the %rdx will be clobbered by "subq %rsi, %rdx"
> below and the %rax should be return value. That's why I copy the %rdi
> twice. memmove() returns the dst pointer. Did I miss something?

I'd forgotten about the (stupid) return value.

I'm pretty sure it is an accident from the original pdp-11
implementation from the days before C had an explicit 'return'
statement.
(The pdp-11 I used ran RSX/11M - so had a Fortran compiler
not a C one.)

	David

-
Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK
Registration No: 1397386 (Wales)

On Fri, Sep 01, 2023 at 02:54:56PM +0000, David Laight wrote:
> I'd forgotten about the (stupid) return value.
> 
> I'm pretty sure it is an accident from the original pdp-11
> implementation from the days before C had an explicit 'return'
> statement.
> (The pdp-11 I used ran RSX/11M - so had a Fortran compiler
> not a C one.)

You're old. I did not exist in that era. And my parents were still young :-)

-- 
Ammar Faizi