Re: [PATCH] x86: static_cpu_has_safe: discard dynamic check after init

[H. Peter Anvin]

On January 18, 2016 5:10:26 PM PST, Borislav Petkov <bp@xxxxxxx> wrote:
>On Mon, Jan 18, 2016 at 07:51:07PM +0100, Borislav Petkov wrote:
>>   [45] .static_cpu_has   PROGBITS         ffffffff97aa655b  16ea655b
>>        00000000000002df  0000000000000000  AX       0     0     1
>
>Ok, staring at this section was wrong. I went and looked at
>the .s file and now it clicked: gcc inlines those calls to
>__static_cpu_has_safe of the dynamic jump target:
>
>        .loc 4 538 0
>        movl    $125, %edi      #,
>        call    __static_cpu_has_safe   #
>.LBE885:
>.LBE886:
>.LBE914:
>        .loc 1 240 0
>        testb   %al, %al        # D.30157
>        je      .L150   #,
>        jmp     .L151   #
>.L152:
>.LBB915:
>.LBB909:
>.LBB905:
>.LBB893:
>.LBB892:
>        .loc 4 538 0
>        movl    $154, %edi      #,
>        call    __static_cpu_has_safe   #
>.LBE892:
>.LBE893:
>.LBE905:
>        .loc 7 431 0
>        testb   %al, %al        # D.30157
>        jne     .L154   #,
>
>which turn into:
>
> 751:   bf 7d 00 00 00          mov    $0x7d,%edi
> 756:   e8 00 00 00 00          callq  75b <fpu__copy+0xab>
> 75b:   84 c0                   test   %al,%al
> 75d:   74 a3                   je     702 <fpu__copy+0x52>
> 75f:   eb 90                   jmp    6f1 <fpu__copy+0x41>
> 761:   bf 9a 00 00 00          mov    $0x9a,%edi
> 766:   e8 00 00 00 00          callq  76b <fpu__copy+0xbb>
> 76b:   84 c0                   test   %al,%al
>
>That's like 28 bytes in this particular case, which is 14 per call
>site,
>on average.
>
>So grepping through my tailored vmlinux, it has 35 entries with
>X86_FEATURE_ALWAYS in .altinstructions X 14 = 490 bytes.
>
>So yeah, we probably should do this, allyesconfig should give more
>savings.
>
>One thing I'd probably do differently is not call the throwaway section
>.static_cpu_has but something like .altinstr_temporary or so and put it
>after the replacement insns:
>
>.altinstr_replacement : AT(ADDR(.altinstr_replacement) - LOAD_OFFSET) {
>                *(.altinstr_replacement)
>		*(.altinstr_temporary)
>        }
>
>so that we know those instructions belong to the alternatives
>mechanism.
>They'll get discared too, of course.
>
>I could just as well be talking a lot of crap, it is waay too late
>here.

Why the f do we call a subroutine for what amounts to a single bt or test instruction?
-- 
Sent from my Android device with K-9 Mail. Please excuse brevity and formatting.