Re: [PATCH] x86emul: Cache the amd_like() output on x86_emulate() entry

["Alejandro Vallejo" <alejandro.garciavallejo@amd.com>]

On Tue Sep 22, 2026 at 3:34 PM CEST, Jan Beulich wrote:
> On 21.09.2026 11:25, Alejandro Vallejo wrote:
> > The current code instantiates amd_like() way too many times, leading to
> > codegen explosion. Unconditionally call it early on, and use that
> > variable everywhere. This shrinks the emulator by ~3KiB.
> > 
> > No a functional change.
> > 
> > Signed-off-by: Alejandro Vallejo <alejandro.garciavallejo@amd.com>
> > ---
> > pipeline: https://gitlab.com/xen-project/people/agvallejo/xen/-/pipelines/2860748404
> >           (it's red because of an caching issue and personal branch name
> >            conventions it's just arm. x86 passes in full)
> > 
> > bloat-o-meter before-after patch.
> > 
> > add/remove: 0/0 grow/shrink: 0/1 up/down: 0/-3366 (-3366)
> > Function                                     old     new   delta
> > x86_emulate                               209351  205985   -3366
> > Total: Before=3617694, After=3614328, chg -0.09%
>
> That's surprisingly big a difference, considering how simple amd_like() is.
> I'm counting 24 instances, so the savings would be well over 100 bytes per
> instance.

I have some more interesting results. I can very much repro at that
commit with GCC 13 with debug=y, but ALSO the variable must be set at
the tail of all other locals. I never noticed the gain was gone after
moving where it is in this patch.

Clearly I misattributed the origin of the shrinkage. In retrospect, it
turns out to be an unrelated matter. I _THINK_ register pressure is
somehow precluding the inlining of a helper. The problem is gone in
debug=n where CSE ensures all is good.

If I find a way of shrinking debug builds in a way that's more resilient
to compiler specifics I'll push that as a separate patch, but for now
I'll just drop this. Clearly it doesn't matter here for a realistic
build.

Apologies for the noise. Codegen is hard :)

Alejandro