Re: [RFC PATCH 00/12] Support vector and more extended registers in perf

[Peter Zijlstra <peterz@infradead.org>]

On Fri, Jun 13, 2025 at 06:49:31AM -0700, kan.liang@linux.intel.com wrote:
> From: Kan Liang <kan.liang@linux.intel.com>
> 
> Starting from the Intel Ice Lake, the XMM registers can be collected in
> a PEBS record. More registers, e.g., YMM, ZMM, OPMASK, SPP and APX, will
> be added in the upcoming Architecture PEBS as well. But it requires the
> hardware support.
> 
> The patch set provides a software solution to mitigate the hardware
> requirement. It utilizes the XSAVES command to retrieve the requested
> registers in the overflow handler. The feature isn't limited to the PEBS
> event or specific platforms anymore.
> The hardware solution (if available) is still preferred, since it has
> low overhead (especially with the large PEBS) and is more accurate.
> 
> In theory, the solution should work for all X86 platforms. But I only
> have newer Inter platforms to test. The patch set only enable the
> feature for Intel Ice Lake and later platforms.
> 
> Open:
> The new registers include YMM, ZMM, OPMASK, SSP, and APX.
> The sample_regs_user/intr has run out. A new field in the
> struct perf_event_attr is required for the registers.
> There could be several options as below for the new field.
> 
> - Follow a similar format to XSAVES. Introduce the below fields to store
>   the bitmap of the registers.
>   struct perf_event_attr {
>         ...
>         __u64   sample_ext_regs_intr[2];
>         __u64   sample_ext_regs_user[2];
>         ...
>   }
>   Includes YMMH (16 bits), APX (16 bits), OPMASK (8 bits),
>            ZMMH0-15 (16 bits), H16ZMM (16 bits), SSP
>   For example, if a user wants YMM8, the perf tool needs to set the
>   corresponding bits of XMM8 and YMMH8, and reconstruct the result.
>   The method is similar to the existing method for
>   sample_regs_user/intr, and match the XSAVES format.
>   The kernel doesn't need to do extra configuration and reconstruction.
>   It's implemented in the patch set.
> 
> - Similar to the above method. But the fields are the bitmap of the
>   complete registers, E.g., YMM (16 bits), APX (16 bits),
>   OPMASK (8 bits), ZMM (32 bits), SSP.
>   The kernel needs to do extra configuration and reconstruction,
>   which may brings extra overhead.
> 
> - Combine the XMM, YMM, and ZMM. So all the registers can be put into
>   one u64 field.
>         ...
>         union {
>                 __u64 sample_ext_regs_intr;   //sample_ext_regs_user is simiar
>                 struct {
>                         __u32 vector_bitmap;
>                         __u32 vector_type   : 3, //0b001 XMM 0b010 YMM 0b100 ZMM
>                               apx_bitmap    : 16,
>                               opmask_bitmap : 8,
>                               ssp_bitmap    : 1,
>                               reserved      : 4,
> 
>                 };
>         ...
>   For example, if the YMM8-15 is required,
>   vector_bitmap: 0x0000ff00
>   vector_type: 0x2
>   This method can save two __u64 in the struct perf_event_attr.
>   But it's not straightforward since it mixes the type and bitmap.
>   The kernel also needs to do extra configuration and reconstruction.
> 
> Please let me know if there are more ideas.

https://lkml.kernel.org/r/20250416155327.GD17910@noisy.programming.kicks-ass.net

comes to mind. Combine that with a rule that reclaims the XMM register
space from perf_event_x86_regs when sample_simd_reg_words != 0, and then
we can put APX and SPP there.