Re: [PATCHv8 02/17] x86/asm: Introduce inline memcpy and memset

[David Laight <david.laight.linux@gmail.com>]

On Thu, 3 Jul 2025 15:33:16 +0200
Vegard Nossum <vegard.nossum@oracle.com> wrote:

> On 03/07/2025 14:15, David Laight wrote:
> > On Thu, 3 Jul 2025 13:39:57 +0300
> > "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com> wrote:  
> >> On Thu, Jul 03, 2025 at 09:44:17AM +0100, David Laight wrote:  
> >>> On Tue,  1 Jul 2025 12:58:31 +0300
> >>> "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com> wrote:  
> >>>> diff --git a/arch/x86/lib/clear_page_64.S b/arch/x86/lib/clear_page_64.S
> >>>> index a508e4a8c66a..47b613690f84 100644
> >>>> --- a/arch/x86/lib/clear_page_64.S
> >>>> +++ b/arch/x86/lib/clear_page_64.S
> >>>> @@ -55,17 +55,26 @@ SYM_FUNC_END(clear_page_erms)
> >>>>   EXPORT_SYMBOL_GPL(clear_page_erms)
> >>>>   
> >>>>   /*
> >>>> - * Default clear user-space.
> >>>> + * Default memset.
> >>>>    * Input:
> >>>>    * rdi destination
> >>>> + * rsi scratch
> >>>>    * rcx count
> >>>> - * rax is zero
> >>>> + * al is value
> >>>>    *
> >>>>    * Output:
> >>>>    * rcx: uncleared bytes or 0 if successful.
> >>>> + * rdx: clobbered
> >>>>    */
> >>>>   SYM_FUNC_START(rep_stos_alternative)
> >>>>   	ANNOTATE_NOENDBR
> >>>> +
> >>>> +	movzbq %al, %rsi
> >>>> +	movabs $0x0101010101010101, %rax
> >>>> +
> >>>> +	/* RDX:RAX = RAX * RSI */
> >>>> +	mulq %rsi  
> >>>
> >>> NAK - you can't do that here.
> >>> Neither %rsi nor %rdx can be trashed.
> >>> The function has a very explicit calling convention.  
> 
> That's why we have the clobbers... see below
> 
> >> What calling convention? We change the only caller to confirm to this.  
> > 
> > The one that is implicit in:
> >   
> >>>> +	asm volatile("1:\n\t"
> >>>> +		     ALT_64("rep stosb",
> >>>> +			    "call rep_stos_alternative", ALT_NOT(X86_FEATURE_FSRM))
> >>>> +		     "2:\n\t"
> >>>> +		     _ASM_EXTABLE_UA(1b, 2b)
> >>>> +		     : "+c" (len), "+D" (addr), ASM_CALL_CONSTRAINT
> >>>> +		     : "a" ((uint8_t)v)  
> > 
> > The called function is only allowed to change the registers that
> > 'rep stosb' uses - except it can access (but not change)
> > all of %rax - not just %al.
> > 
> > See: https://godbolt.org/z/3fnrT3x9r
> > In particular note that 'do_mset' must not change %rax.
> > 
> > This is very specific and is done so that the compiler can use
> > all the registers.  
> 
> I feel like you trimmed off the clobbers from the asm() in the context
> above. For reference, it is:
> 
> +		     : "memory", _ASM_SI, _ASM_DX);

I'm sure they weren't there...

Enough clobbers will 'un-break' it - but that isn't the point.
Linux will reject the patch if he reads it.
The whole point about the function is that it is as direct a replacement
for 'rep stos/movsb' as possible.  

> 
> I'm not saying this can't be optimized, but that doesn't seem to be your
> complaint -- you say "the called function is only allowed to change
> ...", but this is not true when we have the clobbers, right?

You can't change %rax either - not without a clobber.

Oh, and even with your version you only clobbers for %rax and %rdx.
There is no need to use both %rsi and %rdx.

The performance is a different problem.
And the extra clobbers are likely to matter.
x86 really doesn't have many registers.

	David

> 
> This is exactly what I fixed with my v7 fixlet to this patch:
> 
> https://lore.kernel.org/all/1b96b0ca-5c14-4271-86c1-c305bf052b16@oracle.com/
> 
> 
> Vegard