Re: [PATCH 3/4] x86: open-code register save/restore in trace_hardirqs thunks

[Andy Lutomirski <luto@amacapital.net>]

On Sat, Jan 10, 2015 at 12:14 PM, Denys Vlasenko
<vda.linux@googlemail.com> wrote:
> On Sat, Jan 10, 2015 at 3:23 PM, Borislav Petkov <bp@alien8.de> wrote:
>> Bah, I see it. This nasty '$' gets forgotten a lot, maybe we should have
>> a check for that in some scripts :-)
>>
>> Here's the fix:
>>
>> ---
>> Index: b/arch/x86/lib/thunk_64.S
>> ===================================================================
>> --- a/arch/x86/lib/thunk_64.S   2015-01-10 15:18:04.418737613 +0100
>> +++ b/arch/x86/lib/thunk_64.S   2015-01-10 15:17:18.882736556 +0100
>> @@ -67,7 +67,7 @@ restore:
>>         movq_cfi_restore 6*8, rdx
>>         movq_cfi_restore 7*8, rsi
>>         movq_cfi_restore 8*8, rdi
>> -       addq 9*8, %rsp
>> +       addq $9*8, %rsp
>>         CFI_ADJUST_CFA_OFFSET -9*8
>>         ret
>
> Thanks!
>
> After I've seen the disassembly I myself posted, I can't help but wonder
> why we use 5-byte instructions to store and load regs on stack when
> pushes and pops are 1 or 2-byte long.
>

I asked this once, and someone told me that push/pop has lower
throughput.  I find this surprising.

--Andy

> Especially that 32-bit code *does* use push/pops.
>
> Can you test the attached patch with your kvm guest testcase?

Tt could be worth adding a macro along the lines of pushq_cfi_save
that does the pushq_cfi and the CFI_REL_OFFSET.

--Andy

-- 
Andy Lutomirski
AMA Capital Management, LLC