Received-SPF: pass (google.com: domain of linux-kernel-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: <20220324073039.140946-1-ammarfaizi2@gnuweeb.org>
 <20220324073039.140946-5-ammarfaizi2@gnuweeb.org> <20220324075728.GC18586@1wt.eu>
In-Reply-To: <20220324075728.GC18586@1wt.eu>
From:   Alviro Iskandar Setiawan <alviro.iskandar@gnuweeb.org>
Date:   Thu, 24 Mar 2022 15:33:57 +0700
Message-ID: <CAOG64qMwKYHLrUVro1gFhYqHvm8wq5DUdO7QfK5gG2TKhfnNhA@mail.gmail.com>
Subject: Re: [PATCH v1 04/11] tools/nolibc: x86-64: Use appropriate register
 constraints if exist
To:     Willy Tarreau <w@1wt.eu>
Cc:     Ammar Faizi <ammarfaizi2@gnuweeb.org>,
        "Paul E. McKenney" <paulmck@kernel.org>,
        Nugraha <richiisei@gmail.com>,
        Linux Kernel Mailing List <linux-kernel@vger.kernel.org>,
        "GNU/Weeb Mailing List" <gwml@vger.gnuweeb.org>,
        David Laight <David.Laight@aculab.com>
Content-Type: text/plain; charset="UTF-8"
Precedence: bulk

On Thu, Mar 24, 2022 at 2:57 PM Willy Tarreau <w@1wt.eu> wrote:
> On Thu, Mar 24, 2022 at 02:30:32PM +0700, Ammar Faizi wrote:
> > Use appropriate register constraints if exist. Don't use register
> > variables for all inputs.
> >
> > Register variables with "r" constraint should be used when we need to
> > pass data through a specific register to extended inline assembly that
> > doesn't have a specific register constraint associated with it (anything
> > outside %rax, %rbx, %rcx, %rdx, %rsi, %rdi).
> >
> > It also simplifies the macro definition.
>
> I'm a bit bothered by this one because I went the exact opposite route
> in the early design precisely because I found that the current one was
> simpler. [...]
[...]
> I'd say that if there is any technical benefit in doing this (occasional
> code improvement or better support for older or exotic compilers), I'd say
> "ok go for it", but if it's only a matter of taste, I'm not convinced at
> all and am rather seeing this as a regression. Now if there's rough
> consensus around this approach I'll abide, but then I'd request that other
> archs are adapted as well so that we don't keep a different approach only
> for these two ones.

I don't see any technical benefit for x86-64, so I don't think there
is a need in doing this. Though I personally prefer to use register
constraints if they exist instead of register variables for everything
(oh yeah, matter of taste since I don't have any technical argument to
say it's better respecting the resulting codegen). The only real issue
is for the syscall6() implementation on i386 as we've been bitten by a
real compiler issue. In short, I am neutral on this change.

Regards~~
-- Viro