Re: [PATCH V3 01/23] perf/x86: Support outputting XMM registers

["Liang, Kan" <kan.liang@linux.intel.com>]

On 3/23/2019 5:56 AM, Peter Zijlstra wrote:
> On Fri, Mar 22, 2019 at 10:22:50AM -0700, Andi Kleen wrote:
>>>> diff --git a/arch/x86/include/uapi/asm/perf_regs.h b/arch/x86/include/uapi/asm/perf_regs.h
>>>> index f3329cabce5c..b33995313d17 100644
>>>> --- a/arch/x86/include/uapi/asm/perf_regs.h
>>>> +++ b/arch/x86/include/uapi/asm/perf_regs.h
>>>> @@ -28,7 +28,29 @@ enum perf_event_x86_regs {
>>>>   	PERF_REG_X86_R14,
>>>>   	PERF_REG_X86_R15,
>>>>   
>>>> -	PERF_REG_X86_32_MAX = PERF_REG_X86_GS + 1,
>>>> -	PERF_REG_X86_64_MAX = PERF_REG_X86_R15 + 1,
>>>
>>> So this changes UAPI visible symbols... did we think about that?
>>
>> Should be fine. Old programs won't use the new bits,
>> and it just uses not yet used bits.
> 
> Old programs (that used the above symbols) will now fail to compile.
> Even if they won't use the new bits, that seems like a bad thing.
>

Yes, other programs which use the PERF_REG_GPR_X86_32/64_MAX symbols 
should be broken.
I think the new name PERF_REG_GPR_X86_32/64_MAX are more accurate. So I 
will keep both names in V4, and add comments for the old names.

/*
  * These names are deprecated, please use new names as below to instead.
  *     PERF_REG_GPR_X86_32_MAX
  *     PERF_REG_GPR_X86_64_MAX
  */
PERF_REG_X86_32_MAX = PERF_REG_X86_GS + 1,
PERF_REG_X86_64_MAX = PERF_REG_X86_R15 + 1,


>>>> +	/* These all need two bits set because they are 128bit */
>>>> +	PERF_REG_X86_XMM0  = 32,
>>>> +	PERF_REG_X86_XMM1  = 34,
>>>> +	PERF_REG_X86_XMM2  = 36,
>>>> +	PERF_REG_X86_XMM3  = 38,
>>>> +	PERF_REG_X86_XMM4  = 40,
>>>> +	PERF_REG_X86_XMM5  = 42,
>>>> +	PERF_REG_X86_XMM6  = 44,
>>>> +	PERF_REG_X86_XMM7  = 46,
>>>> +	PERF_REG_X86_XMM8  = 48,
>>>> +	PERF_REG_X86_XMM9  = 50,
>>>> +	PERF_REG_X86_XMM10 = 52,
>>>> +	PERF_REG_X86_XMM11 = 54,
>>>> +	PERF_REG_X86_XMM12 = 56,
>>>> +	PERF_REG_X86_XMM13 = 58,
>>>> +	PERF_REG_X86_XMM14 = 60,
>>>> +	PERF_REG_X86_XMM15 = 62,
>>>> +
>>>> +	/* This does not include the XMMX registers */
>>>> +	PERF_REG_GPR_X86_32_MAX = PERF_REG_X86_GS + 1,
>>>> +	PERF_REG_GPR_X86_64_MAX = PERF_REG_X86_R15 + 1,
>>>> +
>>>> +	/* All registers include the XMMX registers */
>>>> +	PERF_REG_X86_MAX = PERF_REG_X86_XMM15 + 2,
>>>>   };
>>>>   #endif /* _ASM_X86_PERF_REGS_H */
>>>
>>> Also, what happens if we run a 32bit kernel or 32bit compat task?
>>>
>>> Then the register dump will report PERF_SAMPLE_REGS_ABI_32, should we
>>> then still interpret the XMM registers as 2x64bit?
>>
>> Yes XMM registers are 128bit in 32bit mode too.
>>
>>>
>>> Are they still at the same offset?
>>
>> Yes.
> 
> I think that is broken.. perf_prepare_sample() does:
> 
>   size += hweight(mask) * sizeof(u64);

It does size += hweight64(mask) * sizeof(u64);

> 
> And since 32bits will not have r8-r15 set, the XMM registers will shift
> forward no?
>

I tried a 32bits kernel, but I didn't observe any issue.

The index of XMM registers always start from 32. That's hard coded.

To double check, I also dumped the mask value in perf_prepare_sample().
With command "perf record -e cycles:p -IXMM0,IXMM1 sleep 1", the mask
value is 0xf00000000, hweight64(mask) returns 4. That is expected.

Is there anything I missed?

>>> Do we need additional PERF_SAMPLE_REGS_ABI_* definitions for this?
>>
>> I don't think so.
> 
> because....?
> 

I didn't observe any broken on 32bit. I think we don't need ABI to 
distinguish the XMM registers.

Thanks,
Kan