Re: [PATCH 1/3 gnumach] apic: Use cpuid to read the apic id for speed

[Luca]

Il 11/08/23 10:34, Damien Zammit ha scritto:> diff --git 
a/i386/i386/cpu_number.h b/i386/i386/cpu_number.h
> index c00896e8..df086370 100644
> --- a/i386/i386/cpu_number.h
> +++ b/i386/i386/cpu_number.h
> @@ -39,12 +39,30 @@
>   #define       CX(addr, reg)   addr(,reg,8)
>   #endif
>   
> -#define	CPU_NUMBER(reg)	\
> +#define        CPU_NUMBER_NO_STACK(reg)        \
>         movl    %cs:lapic, reg          ;\
>         movl    %cs:APIC_ID(reg), reg   ;\
>         shrl    $24, reg                ;\
>         movl    %cs:CX(cpu_id_lut, reg), reg    ;\
>   
> +/* Never call CPU_NUMBER(%esi) */
> +#define CPU_NUMBER(reg)                \
> +       pushl   %esi            ;\
> +       pushl   %eax            ;\
> +       pushl   %ebx            ;\
> +       pushl   %ecx            ;\
> +       pushl   %edx            ;\
> +       movl    $1, %eax        ;\
> +       cpuid                   ;\
> +       shrl    $24, %ebx       ;\
> +       movl    %cs:CX(cpu_id_lut, %ebx), %esi  ;\
> +       popl    %edx            ;\
> +       popl    %ecx            ;\
> +       popl    %ebx            ;\
> +       popl    %eax            ;\
> +       movl    %esi, reg       ;\
> +       popl    %esi            ;\
> +

How much faster is this? did you measure also on hw? I would use RDTSC 
to do it, I don't know if there are better ways.
Interestingly, Linux uses either LSL or RDPID to obtain the same value, 
I wonder if they could be an easy alternative.

Also, could it be that there is a particular use of it that is much more 
frequent than the others, and that might benefit from further optimizations?


Luca