Received-SPF: pass (google.com: domain of linux-crypto-owner@vger.kernel.org designates 23.128.96.18 as permitted sender) client-ip=23.128.96.18;
MIME-Version: 1.0
References: <20210514100106.3404011-1-arnd@kernel.org> <20210514100106.3404011-8-arnd@kernel.org>
 <YKLlyQnR+3uW4ETD@gmail.com> <CAK8P3a0iqe5V6uvaW+Eo0qiwzvyUVavVEfZGwXh4s8ad+0RdCg@mail.gmail.com>
 <CAHk-=wjjo+F8HVkq3eLg+=7hjZPF5mkA4JbgAU8FGE_oAw2MEg@mail.gmail.com>
 <CAK8P3a3hbts4k+rrfnE8Z78ezCaME0UVgwqkdLW5NOps2YHUQQ@mail.gmail.com> <CAHk-=wjuoGyxDhAF8SsrTkN0-YfCx7E6jUN3ikC_tn2AKWTTsA@mail.gmail.com>
In-Reply-To: <CAHk-=wjuoGyxDhAF8SsrTkN0-YfCx7E6jUN3ikC_tn2AKWTTsA@mail.gmail.com>
From:   "Jason A. Donenfeld" <Jason@zx2c4.com>
Date:   Tue, 18 May 2021 20:09:27 +0200
Message-ID: <CAHmME9otB5Wwxp7H8bR_i2uH2esEMvoBMC8uEXBMH9p0q1s6Bw@mail.gmail.com>
Subject: Re: [PATCH v2 07/13] asm-generic: unaligned always use struct helpers
To:     Linus Torvalds <torvalds@linux-foundation.org>
Cc:     Arnd Bergmann <arnd@kernel.org>,
        Eric Biggers <ebiggers@kernel.org>,
        linux-arch <linux-arch@vger.kernel.org>,
        Vineet Gupta <vgupta@synopsys.com>,
        Russell King <linux@armlinux.org.uk>,
        Herbert Xu <herbert@gondor.apana.org.au>,
        "David S. Miller" <davem@davemloft.net>,
        Thomas Bogendoerfer <tsbogend@alpha.franken.de>,
        Linux ARM <linux-arm-kernel@lists.infradead.org>,
        Linux Kernel Mailing List <linux-kernel@vger.kernel.org>,
        "open list:HARDWARE RANDOM NUMBER GENERATOR CORE" 
        <linux-crypto@vger.kernel.org>,
        "open list:BROADCOM NVRAM DRIVER" <linux-mips@vger.kernel.org>
Content-Type: text/plain; charset="UTF-8"
Precedence: bulk

Hi Linus,

On Tue, May 18, 2021 at 6:12 PM Linus Torvalds
<torvalds@linux-foundation.org> wrote:
> I'm actually surprised wireguard would use -O3. Yes, performance is
> important. But for wireguard, correctness is certainly important too.
> Maybe Jason isn't aware of just how bad gcc -O3 has historically been?
> Jason? How big of a deal is that -O3 for wireguard wrt the normal -O2?
> There are known buggy gcc versions that aren't ancient.

My impression has always been that O3 might sometimes generate slower
code, but not that it generates buggy code so commonly. Thanks for
letting me know.

I have a habit of compulsively run IDA Pro after making changes (brain
damage from too many years as a "security person" or something), to
see what the compiler did, and I've just been doing that with O3 since
the beginning of the project, so that's what I wound up optimizing
for. Or sometimes I'll work little things out in Godbolt's compiler
explorer. It's not like it matters much most of the time, but
sometimes I enjoy the golf. Anyway, I've never noticed it producing
any clearly wrong code compared to O2. But I'm obviously not testing
on all compilers or on all architectures. So if you think there's
danger lurking somewhere, it seems reasonable to change that to O2.

Comparing gcc 11's output between O2 and O3, it looks like the primary
difference is that the constant propagation is much less aggressive
with O2, and less inlining in general also means that some stores and
loads to local variables across static function calls aren't being
coalesced. A few null checks are removed too, where the compiler can
prove them away.

So while I've never seen issues with that code under O3, I don't see a
super compelling speed up anywhere either, but rather a bunch of
places that may or may not be theoretically faster or slower on some
system, maybe. I can queue up a patch for the next wireguard series I
send to Dave.

Jason