[PATCH] x86: Remove AMX-TRANSPOSE support

[Jiang, Haochen]

> From: Jan Beulich <jbeulich@suse.com>
> Sent: Friday, October 10, 2025 3:45 PM
> 
> On 10.10.2025 08:20, Haochen Jiang wrote:
> > In ISE059, AMX-TRANSPOSE is removed. It will never happen to any
> > future hardware since x86 Advisory Group is aware of the removal.
> > Thus, We decide to remove the support to ease maintainence effort.
> >
> > Due to the support being in Binutils 2.44/2.45, we should make sure
> > no one use the same encoding in the future or it will conflict
> > between versions. I added comments in the patch in case someone
> > encounter that.
> >
> > Discussions are welcomed for the removal.

Combining the threads. Christian have mentioned that in nasm they chose
not to remove them:

"
for cross-reference: NASM retains AMX-TRANSPOSE support

https://lists.nasm.us/archives/nasm-commits/2025-October/001159.html
"

I am ok with both side. But I still want to list a reason why we could
remove it.

> 
> I'm generally unconvinced of removing things that were once specified,
> and clearly for a reason. AVX10/256 had sensible reasons to exist, and
> transposing matrices is something one can't - afaict - easily
> synthesize from other insns (say, no more than three). Yet use of
> transposed matrices does occur in (mathematical) practice. Hence I find
> it somewhat hard to believe that "it will never happen to any future
> hardware".

Today the x86 Advisory Group just mentioned the future of AMX
as ACE:

https://www.amd.com/en/blogs/2025/amd-and-intel-celebrate-first-anniversary-of-x86-ecosys.html

Since we have a joint opinion at least between AMD and Intel for
matrixes, AMX-TRANSPOSE is really hard to come back.

But as Christian said, we also have other x86 vendors, if we are
concerned on that, we should keep that and not enabling it by
default instead.

> 
> That said, gradually I'm more concerned of the removal from the
> disassembler than that from the assembler. (And just to clarify: I
> think the compiler is in a different position here. Much like e.g. MPX
> and Phi support was removed there, we already made clear that we intend
> to keep it in binutils. Of course that's not exactly an apples-to-
> apples comparison, because there production hardware exists with the
> features available.)
> 
> Further, if we decided to retain support, perhaps bafcf0823c1a
> ("x86/APX: drop AMX-TRANSPOSE promoted insns") would then probably
> also want undoing.

Definitely we should undo that if we want to retain the support.

Thx,
Haochen