Re: [Patch v3 16/22] perf/core: Support to capture higher width vector registers

["Mi, Dapeng" <dapeng1.mi@linux.intel.com>]

On 4/16/2025 11:53 PM, Peter Zijlstra wrote:
> On Wed, Apr 16, 2025 at 02:42:12PM +0800, Mi, Dapeng wrote:
>
>> Just think twice, using bitmap to represent these extended registers indeed
>> wastes bits and is hard to extend, there could be much much more vector
>> registers if considering AMX.
> *Groan* so AMX should never have been register state :-(
>
>
>> Considering different arch/HW may support different number vector register,
>> like platform A supports 8 XMM registers and 8 YMM registers, but platform
>> B only supports 16 XMM registers, a better way to represent these vector
>> registers may add two fields, one is a bitmap which represents which kinds
>> of vector registers needs to be captures. The other field could be a u16
>> array which represents the corresponding register length of each kind of
>> vector register. It may look like this.
>>
>> #define    PERF_SAMPLE_EXT_REGS_XMM    BIT(0)
>> #define    PERF_SAMPLE_EXT_REGS_YMM    BIT(1)
>> #define    PERF_SAMPLE_EXT_REGS_ZMM    BIT(2)
>>     __u32    sample_regs_intr_ext;
>>     __u16    sample_regs_intr_ext_len[4];
>>     __u32    sample_regs_user_ext;
>>     __u16    sample_regs_user_ext_len[4];
>>
>>
>> Peter, how do you think this? Thanks.
> I'm not entirely sure I understand.
>
> How about something like:
>
> 	__u16 sample_simd_reg_words;
> 	__u64 sample_simd_reg_intr;
> 	__u64 sample_simd_reg_user;

If only considering x86 XMM/YMM/ZMM registers, it should be enough since
higher width vector registers always contain the lower width vector
registers on x86 platforms, but I'm not sure if we can have such assumption
for other archs. If not, then it's not enough since user may hope to sample
multiple vector registers with different width at the same time.
Furthermore, considering there could be more other registers like APX
registers need to be supported in the future, we'd better define a more
generic and easily extended interface. That's why I suggest to add a bitmap
like above"sample_regs_intr_ext" which can represent multiple kinds of
registers simultaneously.


>
> Then the simd_reg_words tell us how many (quad) words per register (8 for
> 512) and simd_reg_{intr,user} are a simple bitmap, one bit per actual
> simd reg.
>
> So then all of XMM would be:
>
>   words = 2;
>   intr = user = 0xFFFF;
>
> (16 regs, 128 wide)
>
> Whereas ZMM would be:
>
>   words = 8
>   intr = user = 0xFFFFFFFF;
>
> (32 regs, 512 wide)
>
>
> Would this be sufficient? Possibly we can split the words thing into two
> __u8, but does it make sense to ask for different vector width for
> intr and user ?

Yes, we need it. Users may need to sample interrupt registers and user
space registers simultaneously although it sounds a little bit weird.