Received-SPF: pass (google.com: domain of linux-wireless-owner@vger.kernel.org designates 2620:137:e000::1:20 as permitted sender) client-ip=2620:137:e000::1:20;
MIME-Version: 1.0
References: <20230329-iwlwifi-ptp-avoid-64-bit-div-v1-1-ad8db8d66bc2@kernel.org>
 <9058a032c177e9b04adbf944ad34c5ed8090d9d6.camel@sipsolutions.net>
In-Reply-To: <9058a032c177e9b04adbf944ad34c5ed8090d9d6.camel@sipsolutions.net>
From:   Nick Desaulniers <ndesaulniers@google.com>
Date:   Wed, 29 Mar 2023 10:30:51 -0700
Message-ID: <CAKwvOdnQ9feXGYV2CUyVwg-FNAOmb4HBmDxMg243v28DzSfLuA@mail.gmail.com>
Subject: Re: [PATCH wireless-next] wifi: iwlwifi: mvm: Avoid 64-bit division
 in iwl_mvm_get_crosstimestamp_fw()
To:     Johannes Berg <johannes@sipsolutions.net>
Cc:     Nathan Chancellor <nathan@kernel.org>, gregory.greenman@intel.com,
        kvalo@kernel.org, trix@redhat.com, avraham.stern@intel.com,
        krishnanand.prabhu@intel.com, linux-wireless@vger.kernel.org,
        netdev@vger.kernel.org, linux-kernel@vger.kernel.org,
        llvm@lists.linux.dev, patches@lists.linux.dev,
        Arnd Bergmann <arnd@arndb.de>,
        "kernelci.org bot" <bot@kernelci.org>,
        Craig Topper <craig.topper@sifive.com>
Content-Type: text/plain; charset="UTF-8"
Content-Transfer-Encoding: quoted-printable
Precedence: bulk

On Wed, Mar 29, 2023 at 10:20=E2=80=AFAM Johannes Berg
<johannes@sipsolutions.net> wrote:
>
> On Wed, 2023-03-29 at 10:05 -0700, Nathan Chancellor wrote:
> >
> > GCC has optimizations for division by a constant that clang does not
> > implement, so this issue is not visible when building with GCC.
>
> Huh yeah, we did 32-bit builds with gcc ...

Right, GCC is better about turning division by double-word constant
into multiplication by reciprocal. Craig has been improving LLVM, but
it seems that some divisors still aren't supported (in this case 100).

>
> > Using div_u64() would resolve this issue, but Arnd points out that this
> > can be quite expensive and the timestamp is being read at nanosecond
> > granularity.
>
> Doesn't matter though, all the calculations are based on just the
> command response from the firmware, which (tries to) take it in a
> synchronised fashion.
>
> So taking more time here would be fine, as far as I can tell.

div_u64() it is then.

>
> > Nick pointed out that the result of this division is being
> > stored to a 32-bit type anyways, so truncate gp2_10ns first then do the
> > division, which elides the need for libcalls.
>
> That loses ~7 top bits though, no? I'd be more worried about that, than
> the time div_u64() takes.

The result is still stored in a u32; there is a loss of precision
regardless of use of div_u64 or open coded binary operator /.  So is
the loss of precision before the division as tolerable as after the
division?

--=20
Thanks,
~Nick Desaulniers