Re: [PATCH v1] x86: Add EVEX optimized memchr family not safe for RTM

[Noah Goldstein <goldstein.w.n@gmail.com>]

On Wed, May 5, 2021 at 11:29 AM H.J. Lu <hjl.tools@gmail.com> wrote:
>
> On Wed, May 5, 2021 at 11:19 AM Noah Goldstein <goldstein.w.n@gmail.com> wrote:
> >
> > On Wed, May 5, 2021 at 1:55 PM H.J. Lu <hjl.tools@gmail.com> wrote:
> > >
> > > On Wed, May 5, 2021 at 9:25 AM Noah Goldstein <goldstein.w.n@gmail.com> wrote:
> > > >
> > > > On Wed, May 5, 2021 at 9:23 AM H.J. Lu <hjl.tools@gmail.com> wrote:
> > > > >
> > > > > On Tue, May 4, 2021 at 4:34 PM Noah Goldstein <goldstein.w.n@gmail.com> wrote:
> > > > > >
> > > > > > No bug.
> > > > > >
> > > > > > This commit adds a new implementation for EVEX memchr that is not safe
> > > > > > for RTM because it uses vzeroupper. The benefit is that by using
> > > > >
> > > > > EVEX memchr won't cause RTM abort if YMM16-YMM31 are used
> > > > > since there is no need to use vzeroupper.  Please remove vzeroupper from
> > > > > EVEX memchr and remove EVEX RTM functions.
> > > >
> > > > That's impossible for this implementation.
> > > >
> > > > The reason ymm0-ymm15 are used is so that we can use vpcmpeq which is
> > > > not encodable with ymm16-ymm31.
> > > >
> > > > This implementation is optimized for CPUs which dont support RTM but
> > > > do support EVEX.
> > > >
> > >
> > > Are you seeing something along the line of Prefer_AVX2_STRCMP:
> >
> > Yes.
> >
> > For atleast some functions I think EVEX + AVX2 is probably the ideal
> > implementation unless you have to worry about RTM. And as of right now
> > there are a fair amount of x86_64 chips out there w/ avx512 but w/o
> > RTM (or intel might fix the issue where vzeroupper aborts transactions
> > in the future)
> >
> > For small values even if EVEX costs an extra instruction (i.e strchr
> > vpxor + vpmin + vpcmp can be replace with vpcmpeq + vpcmp) the
> > overhead of vzeroupper make EVEX perform better. But once the main
> > loop is hit the overhead of vzeroupper isn't really a concern and
> > vpcmpeq isnt really replaceable with vpcmp from a logic perspective
> > (i.e 3x vpcmpeq + vptern isnt doable with vpcmp) and vpcmp is slower
> > for tput and latency. As a little tackon the EVEX instructions are all
> > larger code footprint than AVX2 so keeping instruction length <= 6byte
> > for the DSB isnt really doable.
>
> The only reason to use YMM16-YMM31 is to avoid vzeroupper.   Since
> we are going to issue vzeroupper anyway,  please use YMM0 to YMM15.

Throughout or just in the 4x loop? Using YMM0-YMM15 throughout is a
regression for the small sizes where the overhead of vzeroupper is
meaningful. I.e the strchr example above.

>
> > I havent taken that long of a look at strcmp but would guess it would
> > benefit in a simliar way as memchr from EVEX for sizes [0..160] then
> > AVX2 for the loop (possibly augmented with vptern for 3way reduction)
> >
> > >
> > > commit 1da50d4bda07f04135dca39f40e79fc9eabed1f8
> > > Author: H.J. Lu <hjl.tools@gmail.com>
> > > Date:   Fri Feb 26 05:36:59 2021 -0800
> > >
> > >     x86: Set Prefer_No_VZEROUPPER and add Prefer_AVX2_STRCMP
> > >
> > >     1. Set Prefer_No_VZEROUPPER if RTM is usable to avoid RTM abort triggered
> > >     by VZEROUPPER inside a transactionally executing RTM region.
> > >     2. Since to compare 2 32-byte strings, 256-bit EVEX strcmp requires 2
> > >     loads, 3 VPCMPs and 2 KORDs while AVX2 strcmp requires 1 load, 2 VPCMPEQs,
> > >     1 VPMINU and 1 VPMOVMSKB, AVX2 strcmp is faster than EVEX strcmp.  Add
> > >     Prefer_AVX2_STRCMP to prefer AVX2 strcmp family functions.
> > >
> > >
> > > --
> > > H.J.
>
>
>
> --
> H.J.