Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::3:8 as permitted sender) client-ip=2620:137:e000::3:8;
Date:   Sat, 16 Sep 2023 16:59:32 +0100
From:   Matthew Wilcox <willy@infradead.org>
To:     Linus Torvalds <torvalds@linux-foundation.org>
Cc:     linux-kernel@vger.kernel.org, linux-fsdevel@vger.kernel.org,
        linux-arch@vger.kernel.org, Nicholas Piggin <npiggin@gmail.com>
Subject: Re: [PATCH 08/17] alpha: Implement xor_unlock_is_negative_byte
Message-ID: <ZQXQ5OnAgisxVyKs@casper.infradead.org>
References: <20230915183707.2707298-1-willy@infradead.org>
 <20230915183707.2707298-9-willy@infradead.org>
 <CAHk-=wgBUvM7tc70AAvUw+HHOo6Q=jD4FVheFGDCjNaK3OCEGA@mail.gmail.com>
 <ZQT4/gA4vIa/7H6q@casper.infradead.org>
 <CAHk-=whbj+pVGhJTcQCLhY8KZJNomWOKM=s-GZSpK_G=G4fXEA@mail.gmail.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <CAHk-=whbj+pVGhJTcQCLhY8KZJNomWOKM=s-GZSpK_G=G4fXEA@mail.gmail.com>
Precedence: bulk

On Fri, Sep 15, 2023 at 07:01:14PM -0700, Linus Torvalds wrote:
> On Fri, 15 Sept 2023 at 17:38, Matthew Wilcox <willy@infradead.org> wrote:
> >
> > On Fri, Sep 15, 2023 at 05:27:17PM -0700, Linus Torvalds wrote:
> > > On Fri, 15 Sept 2023 at 11:37, Matthew Wilcox (Oracle)
> > > <willy@infradead.org> wrote:
> > > >
> > > > +       "1:     ldl_l %0,%4\n"
> > > > +       "       xor %0,%3,%0\n"
> > > > +       "       xor %0,%3,%2\n"
> > > > +       "       stl_c %0,%1\n"
> > >
> > > What an odd thing to do.
> > >
> > > Why don't you just save the old value? That double xor looks all kinds
> > > of strange, and is a data dependency for no good reason that I can
> > > see.
> > >
> > > Why isn't this "ldl_l + mov %0,%2 + xor + stl_c" instead?
> > >
> > > Not that I think alpha matters, but since I was looking through the
> > > series, this just made me go "Whaa?"
> >
> > Well, this is my first time writing Alpha assembler ;-)  I stole this
> > from ATOMIC_OP_RETURN:
> >
> >         "1:     ldl_l %0,%1\n"                                          \
> >         "       " #asm_op " %0,%3,%2\n"                                 \
> >         "       " #asm_op " %0,%3,%0\n"                                 \
> 
> Note how that does "orig" assignment first (ie the '%2" destination is
> the first instruction), unlike your version.

Wow.  I totally missed that I'd transposed those two lines.  I read
it back with the lines in the order that they should have been in.
Every time I read it.  I was wondering why you were talking about a data
dependency, and I just couldn't see it.  With the lines in the order that
they're actually in, it's quite obvious and totally not what I meant.
Of course, it doesn't matter which order they're in from the point of
view of testing the waiters bit since we don't change the waiters bit.

> Does any of this matter? Nope. It's alpha. There's probably a handful
> of machines, and it's maybe one extra cycle. It's really the oddity
> that threw me.

I'll admit to spending far more time on the m68k version of this than
the alpha version ;-)