Re: [PATCH v2 15/15] x86/lib/memcpy_64.S: Convert memcpy to ALTERNATIVE_2 macro

[Ingo Molnar <mingo@kernel.org>]

* Borislav Petkov <bp@alien8.de> wrote:

> From: Borislav Petkov <bp@suse.de>
> 
> Make REP_GOOD variant the default after alternatives have run.
> 
> Signed-off-by: Borislav Petkov <bp@suse.de>
> ---
>  arch/x86/lib/memcpy_64.S | 68 +++++++++++++++---------------------------------
>  1 file changed, 21 insertions(+), 47 deletions(-)
> 
> diff --git a/arch/x86/lib/memcpy_64.S b/arch/x86/lib/memcpy_64.S
> index bbfdacc01760..b046664f5a1c 100644
> --- a/arch/x86/lib/memcpy_64.S
> +++ b/arch/x86/lib/memcpy_64.S
> @@ -1,12 +1,20 @@
>  /* Copyright 2002 Andi Kleen */
>  
>  #include <linux/linkage.h>
> -
>  #include <asm/cpufeature.h>
>  #include <asm/dwarf2.h>
>  #include <asm/alternative-asm.h>
>  
>  /*
> + * We build a jump to memcpy_orig by default which gets NOPped out on
> + * the majority of x86 CPUs which set REP_GOOD. In addition, CPUs which
> + * have the enhanced REP MOVSB/STOSB feature (ERMS), change those NOPs
> + * to a jmp to memcpy_erms which does the REP; MOVSB mem copy.
> + */
> +
> +.weak memcpy
> +
> +/*
>   * memcpy - Copy a memory block.
>   *
>   * Input:
> @@ -17,15 +25,11 @@
>   * Output:
>   * rax original destination
>   */
> +ENTRY(__memcpy)
> +ENTRY(memcpy)
> +	ALTERNATIVE_2 "jmp memcpy_orig", "", X86_FEATURE_REP_GOOD, \
> +		      "jmp memcpy_erms", X86_FEATURE_ERMS
>  
> +ENDPROC(memcpy)
> +ENDPROC(__memcpy)

> +ENTRY(memcpy_erms)
>  	movq %rdi, %rax
>  	movq %rdx, %rcx
>  	rep movsb
>  	ret

Since most CPUs we care about have ERMS, wouldn't it be better to 
patch in the actual memcpy_erms sequence into the primary memcpy() 
function? It's just about 9 bytes AFAICT.

This would remove a jump instruction from the most common memcpy 
variant: worth it!

Thanks,

	Ingo