Re: [PATCH 2/2] x86/bitops: Fix false output register dependency of TZCNT insn

[Uros Bizjak]

On Fri, Mar 28, 2025 at 11:28 PM Ingo Molnar <mingo@xxxxxxxxxx> wrote:
>
>
> * Uros Bizjak <ubizjak@xxxxxxxxx> wrote:
>
> > On Tue, Mar 25, 2025 at 10:43 PM Ingo Molnar <mingo@xxxxxxxxxx> wrote:
> > >
> > >
> > > * Uros Bizjak <ubizjak@xxxxxxxxx> wrote:
> > >
> > > > On Haswell and later Intel processors, the TZCNT instruction appears
> > > > to have a false dependency on the destination register. Even though
> > > > the instruction only writes to it, the instruction will wait until
> > > > destination is ready before executing. This false dependency
> > > > was fixed for Skylake (and later) processors.
> > > >
> > > > Fix false dependency by clearing the destination register first.
> > > >
> > > > The x86_64 defconfig object size increases by 4215 bytes:
> > > >
> > > >           text           data     bss      dec            hex filename
> > > >       27342396        4642999  814852 32800247        1f47df7 vmlinux-old.o
> > > >       27346611        4643015  814852 32804478        1f48e7e vmlinux-new.o
> > >
> > > Yeah, so Skylake was released in 2015, about a decade ago.
> > >
> > > So we'd be making the kernel larger for an unquantified
> > > micro-optimization for CPUs that almost nobody uses anymore.
> > > That's a bad trade-off.
> >
> > Yes, 4.2k seems a bit excessive. OTOH, I'd not say that the issue is
> > a micro-optimization, it is bordering on the hardware bug.
>
> Has this been quantified, and do we really care about the
> micro-performance of ~10-year old CPUs, especially at the
> expense of modern CPUs?

No, although the change would be a one liner now. Without specially
crafted benchmark loops the impact is not noticeable and typical
kernel usage of these instructions is not that sensitive on
destination.

Thanks,
Uros.