Re: [PATCH] arm64: save movk instructions in mov_q when the lower 16|32 bits are all zero

[Ard Biesheuvel]

On Thu, 28 Jul 2022 at 08:26, Jisheng Zhang <jszhang@xxxxxxxxxx> wrote:
>
> On Thu, Jul 28, 2022 at 10:49:02PM +0800, Jisheng Zhang wrote:
> > On Wed, Jul 27, 2022 at 08:15:11AM -0700, Ard Biesheuvel wrote:
> > > On Sat, 9 Jul 2022 at 01:58, Jisheng Zhang <jszhang@xxxxxxxxxx> wrote:
> > > >
> > > > Currently mov_q is used to move a constant into a 64-bit register,
> > > > when the lower 16 or 32bits of the constant are all zero, the mov_q
> > > > emits one or two useless movk instructions. If the mov_q macro is used
> > > > in hot code path, we want to save the movk instructions as much as
> > > > possible. For example, when CONFIG_ARM64_MTE is 'Y' and
> > > > CONFIG_KASAN_HW_TAGS is 'N', the following code in __cpu_setup()
> > > > routine is the pontential optimization target:
> > > >
> > > >         /* set the TCR_EL1 bits */
> > > >         mov_q   x10, TCR_MTE_FLAGS
> > > >
> > > > Before the patch:
> > > >         mov     x10, #0x10000000000000
> > > >         movk    x10, #0x40, lsl #32
> > > >         movk    x10, #0x0, lsl #16
> > > >         movk    x10, #0x0
> > > >
> > > > After the patch:
> > > >         mov     x10, #0x10000000000000
> > > >         movk    x10, #0x40, lsl #32
> > > >
> > > > Signed-off-by: Jisheng Zhang <jszhang@xxxxxxxxxx>
> > >
> > > This is broken for constants that have 0xffff in the top 16 bits, as
> > > in that case, we will emit a MOVN/MOVK/MOVK sequence, and omitting the
> > > MOVKs will set the corresponding field to 0xffff not 0x0.
> >
> > Thanks so much for this hint. I think you are right about the 0xffff in
> > top 16bits case.
> >
>
> the patch breaks below usage case:
> mov_q x0, 0xffffffff00000000
>
> I think the reason is mov_q starts from high bits, if we change the
> macro to start from LSB, then that could solve the breakage. But this
> needs a rewrite of mov_q

No it has nothing to do with that.

The problem is that the use of MOVN changes the implicit value of the
16-bit fields that are left unspecified, and assigning them in a
different order is not going to make any difference.

I don't think we should further complicate mov_q, and I would argue
that the existing optimization (which I added myself) is premature
already: in the grand scheme of things, one or two instructions more
or less are not going to make a difference anyway, given how rarely
this macro is used. And even if any of these occurrences are on a hot
path, it is not a given that shorter sequences of MOVZ/MOVN/MOVK are
going to execute any faster, as the canonical MOVZ/MOVK/MOVK/MOVK
might well decode to fewer uops.

So in summary, let's leave this code be.