Re: [RFC PATCH 06/12] perf: Support extension of sample_regs

[Mark Rutland <mark.rutland@arm.com>]

On Tue, Jun 17, 2025 at 04:06:17PM +0200, Peter Zijlstra wrote:
> On Tue, Jun 17, 2025 at 03:33:33PM +0200, Peter Zijlstra wrote:
> > On Tue, Jun 17, 2025 at 08:14:36PM +0800, Mi, Dapeng wrote:
> > 
> > > > We're going to do a sane SIMD register set with variable width, and
> > > > reclaim the XMM regs from the normal set.
> > > 
> > > Ok, so we need to add two width variables like
> > > sample_ext_regs_words_intr/user,
> > 
> > s/ext/simd/
> > 
> > Not sure it makes sense to have separate vector widths for kernel and
> > user regs, but sure.
> > 
> > > then reuse the XMM regs bitmap to represent the extend regs bitmap.
> > 
> > But its not extended; its the normal bitmap.
> > 
> > > Considering the OPMASK regs and APX
> > > extended GPRs have same bit-width (64 bits), we may have to combine them
> > > into a single bitmap, e.g. bits[15:0] represents R31~R16 and bits[23:16]
> > > represents OPMASK7 ~ OPMASK0. 
> > 
> > Again confused, bits 0:23 are the normal registers (in a lunatic
> > order). The XMM regs are in 32:63 and will be free if the SIMD thing is
> > present.
> > 
> > SPP+APX should definitely go there.
> > 
> > Not sure about OPMASK; those really do belong with the SIMD state. Let
> > me go figure out what ARM and Risc-V look like in more detail.
> 
> So ARM-SVE has 32 vector registers with 16 predicate registers.
> 
> Risc-V Zv seems to only have 32 vector registers; no special purpose
> predicate registers, instead a regular vector register can be used as a
> predicate register.
> 
> PowerPC VSX has 64 vector registers and no predicate registers afaict.
> 
> While reading this, I came across the useful note that predicate
> registers are 1/8-th the length of the vector registers (because the
> minimal element is a byte). So while the current AVX-512 predicate
> registers are indeed 64bits, this would no longer be true for the
> hypothetical AVX-1024 (or even AVX-512 if we allow 4bit elements).
> 
> As such, I don't think we should stick the predicate registers in the
> normal group -- they really are not normal registers and won't fit for
> future extensions.
> 
> This then leaves us two options:
> 
>  - stick the predicate registers in the high bits of the vector register
>    word, or
> 
>  - add an explicit predicate register word.

TBH, I don't think we can handle extended state in a generic way unless
we treat this like a ptrace regset, and delegate the format of each
specific register set to the architecture code.

On arm64, the behaviour is modal (with two different vector lengths for
streaming/non-streaming SVE when SME is implemented), per-task
configurable (with different vector lengths), can differ between
host/guest for KVM, and some of the registers only exist in some
configurations (e.g. the FFR only exists for SME if FA64 is
implemented).

Mark.