Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::3:4 as permitted sender) client-ip=2620:137:e000::3:4;
MIME-Version: 1.0
References: <ZR2U4DLycLT5xFH6@gmail.com> <20231004192404.31733-1-ubizjak@gmail.com>
 <CAHk-=wjuRGzhuETLYDoi4hM6RAxHVL0ptuRb3TH-od+348Y8zA@mail.gmail.com>
In-Reply-To: <CAHk-=wjuRGzhuETLYDoi4hM6RAxHVL0ptuRb3TH-od+348Y8zA@mail.gmail.com>
From:   Uros Bizjak <ubizjak@gmail.com>
Date:   Wed, 4 Oct 2023 22:07:54 +0200
Message-ID: <CAFULd4YRHmQVnwaORm7=7kUs7DYG7SfwdTXAitDt=bxiMU5AoQ@mail.gmail.com>
Subject: Re: [PATCH v2 4/4] x86/percpu: Use C for percpu read/write accessors
To:     Linus Torvalds <torvalds@linux-foundation.org>
Cc:     x86@kernel.org, linux-kernel@vger.kernel.org,
        Andy Lutomirski <luto@kernel.org>,
        Ingo Molnar <mingo@kernel.org>, Nadav Amit <namit@vmware.com>,
        Brian Gerst <brgerst@gmail.com>,
        Denys Vlasenko <dvlasenk@redhat.com>,
        "H . Peter Anvin" <hpa@zytor.com>,
        Peter Zijlstra <peterz@infradead.org>,
        Thomas Gleixner <tglx@linutronix.de>,
        Borislav Petkov <bp@alien8.de>,
        Josh Poimboeuf <jpoimboe@redhat.com>
Content-Type: text/plain; charset="UTF-8"
Content-Transfer-Encoding: quoted-printable
Precedence: bulk

On Wed, Oct 4, 2023 at 9:42=E2=80=AFPM Linus Torvalds
<torvalds@linux-foundation.org> wrote:
>
> Unrelated reaction..
>
> On Wed, 4 Oct 2023 at 12:24, Uros Bizjak <ubizjak@gmail.com> wrote:
> >
> > the code improves from:
> >
> >  65 8b 05 00 00 00 00    mov    %gs:0x0(%rip),%eax
> >  a9 00 00 0f 00          test   $0xf0000,%eax
> >
> > to:
> >
> >  65 f7 05 00 00 00 00    testl  $0xf0000,%gs:0x0(%rip)
> >  00 00 0f 00
>
> Funky.
>
> Why does gcc generate that full-width load from memory, and not demote
> it to a byte test?

It does when LSB is accessed at the same address. For example:

int m;
_Bool foo (void) { return m & 0x0f; }

compiles to:

  0:   f6 05 00 00 00 00 0f    testb  $0xf,0x0(%rip)        # 7 <foo+0x7>

>
> IOW, it should not be
>
>   65 f7 05 00 00 00 00 testl  $0xf0000,%gs:0x0(%rip)
>   00 00 0f 00
>
> after optimizing it, it should be three bytes shorter at
>
>   65 f6 05 00 00 00 00 testb  $0xf,%gs:0x0(%rip)
>   0f
>
> instead (this is "objdump", so it doesn't show that the relocation
> entry has changed by +2 to compensate).
>
> Now, doing the access narrowing is a bad idea for stores (because it
> can cause subsequent loads to have conflicts in the store buffer), but
> for loads it should always be a win to narrow the access.
>
> I wonder why gcc doesn't do it. This is not related to __seg_gs - I
> tried it with regular memory accesses too, and gcc kept those as
> 32-bit accesses too.
>
> And no, the assembler can't optimize that operation either, since I
> think changing the testl to a testb would change the 'P' bit in the
> resulting eflags, so this is a "the compiler could pick a better
> instruction choice" thing.
>
> I'm probably missing some reason why gcc wouldn't do this. But clang
> does seem to do this obvious optimization.

You get a store forwarding stall when you write a bigger operand to
memory and then read part of it, if the smaller part doesn't start at
the same
address.

Uros.