Re: [PATCH] x86: branchless clear_page()

[Brian Gerst]

On Mon, Aug 22, 2016 at 4:52 PM, Alexey Dobriyan <adobriyan@xxxxxxxxx> wrote:
> Apply alternatives at the call site instead of function body.
> Save branch per clean page.
>
> Bonus: tell gcc to not flush whole shebang of registers,
> just RDI, RAX, RCX.
>
> Signed-off-by: Alexey Dobriyan <adobriyan@xxxxxxxxx>
> ---
>
>  arch/x86/include/asm/page_64.h |   16 +++++++++++++++-
>  arch/x86/lib/clear_page_64.S   |   18 ++++++------------
>  2 files changed, 21 insertions(+), 13 deletions(-)
>
> --- a/arch/x86/include/asm/page_64.h
> +++ b/arch/x86/include/asm/page_64.h
> @@ -4,6 +4,7 @@
>  #include <asm/page_64_types.h>
>
>  #ifndef __ASSEMBLY__
> +#include <asm/alternative.h>
>
>  /* duplicated to the one in bootmem.h */
>  extern unsigned long max_pfn;
> @@ -34,7 +35,20 @@ extern unsigned long __phys_addr_symbol(unsigned long);
>  #define pfn_valid(pfn)          ((pfn) < max_pfn)
>  #endif
>
> -void clear_page(void *page);
> +void clear_page_mov(void *page);
> +void clear_page_rep_stosq(void *page);
> +void clear_page_rep_stosb(void *page);
> +static __always_inline void clear_page(void *page)
> +{
> +       alternative_call_2(
> +               clear_page_mov,
> +               clear_page_rep_stosq, X86_FEATURE_REP_GOOD,
> +               clear_page_rep_stosb, X86_FEATURE_ERMS,
> +               "=D" (page),
> +               "0" (page)
> +               : "rax", "rcx", "memory"
> +       );
> +}
>  void copy_page(void *to, void *from);
>
>  #endif /* !__ASSEMBLY__ */
> --- a/arch/x86/lib/clear_page_64.S
> +++ b/arch/x86/lib/clear_page_64.S
> @@ -1,6 +1,4 @@
>  #include <linux/linkage.h>
> -#include <asm/cpufeatures.h>
> -#include <asm/alternative-asm.h>
>
>  /*
>   * Most CPUs support enhanced REP MOVSB/STOSB instructions. It is
> @@ -13,18 +11,14 @@
>   * Zero a page.
>   * %rdi        - page
>   */
> -ENTRY(clear_page)
> -
> -       ALTERNATIVE_2 "jmp clear_page_orig", "", X86_FEATURE_REP_GOOD, \
> -                     "jmp clear_page_c_e", X86_FEATURE_ERMS
> -
> +ENTRY(clear_page_rep_stosq)
>         movl $4096/8,%ecx
>         xorl %eax,%eax
>         rep stosq
>         ret
> -ENDPROC(clear_page)
> +ENDPROC(clear_page_rep_stosq)
>
> -ENTRY(clear_page_orig)
> +ENTRY(clear_page_mov)
>
>         xorl   %eax,%eax
>         movl   $4096/64,%ecx
> @@ -44,11 +38,11 @@ ENTRY(clear_page_orig)
>         jnz     .Lloop
>         nop
>         ret
> -ENDPROC(clear_page_orig)
> +ENDPROC(clear_page_mov)
>
> -ENTRY(clear_page_c_e)
> +ENTRY(clear_page_rep_stosb)
>         movl $4096,%ecx
>         xorl %eax,%eax
>         rep stosb
>         ret
> -ENDPROC(clear_page_c_e)
> +ENDPROC(clear_page_rep_stosb)

I like this idea, but does it make sense to take it a step further and
inline the string instruction alternatives to avoid a call altogether?

Also, 32-bit should be converted to do the same thing as 64-bit.

--
Brian Gerst