Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 23.128.96.18 as permitted sender) client-ip=23.128.96.18;
MIME-Version: 1.0
References: <20210721135926.602840-1-nborisov@suse.com> <CAHk-=whqJKKc9wUacLEkvTzXYfYOUDt=kHKX6Fa8Kb4kQftbbQ@mail.gmail.com>
 <20210721201029.GQ19710@twin.jikos.cz>
In-Reply-To: <20210721201029.GQ19710@twin.jikos.cz>
From:   Linus Torvalds <torvalds@linux-foundation.org>
Date:   Wed, 21 Jul 2021 13:27:56 -0700
Message-ID: <CAHk-=whCygw44p30Pmf+Bt8=LVtmij3_XOxweEA3OQNruhMg+A@mail.gmail.com>
Subject: Re: [PATCH] lib/string: Bring optimized memcmp from glibc
To:     David Sterba <dsterba@suse.cz>,
        Linus Torvalds <torvalds@linux-foundation.org>,
        Nikolay Borisov <nborisov@suse.com>,
        Linux Kernel Mailing List <linux-kernel@vger.kernel.org>,
        Nick Desaulniers <ndesaulniers@google.com>,
        linux-fsdevel <linux-fsdevel@vger.kernel.org>,
        Dave Chinner <david@fromorbit.com>
Content-Type: text/plain; charset="UTF-8"
Precedence: bulk

On Wed, Jul 21, 2021 at 1:13 PM David Sterba <dsterba@suse.cz> wrote:
>
> adding a memcmp_large that compares by native words or u64 could be
> the best option.

Yeah, we could just special-case that one place.

But see the patches I sent out - I think we can get the best of both worlds.

A small and simple memcmp() that is good enough and not the
_completely_ stupid thing we have now.

The second patch I sent out even gets the mutually aligned case right.

Of course, the glibc code also ended up unrolling things a bit, but
honestly, the way it did it was too disgusting for words.

And if it really turns out that the unrolling makes a big difference -
although I doubt it's meaningful with any modern core - I can add a
couple of lines to that simple patch I sent out to do that too.
Without getting the monster that is that glibc code.

Of course, my patch depends on the fact that "get_unaligned()" is
cheap on all CPU's that really matter, and that caches aren't
direct-mapped any more. The glibc code seems to be written for a world
where registers are cheap, unaligned accesses are prohibitively
expensive, and unrolling helps because L1 caches are direct-mapped and
you really want to do chunking to not get silly way conflicts.

If old-style Sparc or MIPS was our primary target, that would be one
thing. But it really isn't.

              Linus