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;
Message-ID: <b1617fe221f137b49b9a2ba9c864e6b85f89de81.camel@xry111.site>
Subject: Re: [PATCH] LoongArch: Drop -ffreestanding from CFLAGS
From:   Xi Ruoyao <xry111@xry111.site>
To:     Nick Desaulniers <ndesaulniers@google.com>,
        Huacai Chen <chenhuacai@kernel.org>
Cc:     WANG Xuerui <kernel@xen0n.name>,
        Nathan Chancellor <nathan@kernel.org>,
        loongarch@lists.linux.dev, llvm@lists.linux.dev,
        linux-kernel@vger.kernel.org, WANG Xuerui <git@xen0n.name>
Date:   Fri, 04 Aug 2023 23:56:07 +0800
In-Reply-To: <CAKwvOdmt+p3KM23hz2BfheoWgEa0G9QKnAFYUP+EL2Ta5LXUUg@mail.gmail.com>
References: <20230804125609.2054719-1-kernel@xen0n.name>
         <CAAhV-H5g-jmeQMhicrqhpStV9n0Z-EZ1dZAsuT2hmOaHki4ucg@mail.gmail.com>
         <CAKwvOdmt+p3KM23hz2BfheoWgEa0G9QKnAFYUP+EL2Ta5LXUUg@mail.gmail.com>
Content-Type: text/plain; charset="UTF-8"
Content-Transfer-Encoding: quoted-printable
User-Agent: Evolution 3.48.4 
MIME-Version: 1.0
Precedence: bulk

On Fri, 2023-08-04 at 08:46 -0700, Nick Desaulniers wrote:
> > Not all processors support unaligned access, so we need the
> > alternative mechanism to select memset/memcpy/memmove implementations.
> > If remove -ffreestanding, the builtin implementation cannot be used on
> > all hardware.
>=20
> That sounds like a compiler bug in that compiler's implementation of
> string.h builtins then; it should default to the safest implementation
> (aligned accesses) until instructed otherwise.=C2=A0 Have you filed a bug
> against your compiler vendor for which compiler you observe that
> behavior from?

AFAIK there is no such bug in GCC.

But GCC indeed has a bug about builtin expansion: it generates really
stupid code for __builtin_memcpy and friends.  See
https://gcc.gnu.org/PR109465.  The bug is fixed for GCC 14, but GCC 14
won't be released soon.

> At the very least, there should be a comment above the addition of
> -ffreestanding justifying why it's being used, probably with a link to
> the above bug report.
>=20
> I would expect either -mcpu or perhaps some other -m flag to guide the
> compiler when it is safe to emit memcpy (and friends) in terms of
> unaligned access or not.

It's controlled by -m{no-,}strict-align.  LoongArch kernel defaults to -
mstrict-align, with this the compiler should not generate unaligned
access.  -mno-strict-align is hidden behind CONFIG_EXPERT (FWIW I
personally dislike the decision to hide it).

--=20
Xi Ruoyao <xry111@xry111.site>
School of Aerospace Science and Technology, Xidian University