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: <20230410115734.93365-1-wangrui@loongson.cn> <b80e083baa49f0443abee66eb888fac920f66e79.camel@xry111.site>
In-Reply-To: <b80e083baa49f0443abee66eb888fac920f66e79.camel@xry111.site>
From:   Rui Wang <wangrui@loongson.cn>
Date:   Mon, 10 Apr 2023 22:22:39 +0800
Message-ID: <CAHirt9ib2LPXMbxb+UCJB1MsBnUuWNLb5+uyqcR5Wiw2fsJi6w@mail.gmail.com>
Subject: Re: [PATCH] LoongArch: Improve memory ops
To:     Xi Ruoyao <xry111@xry111.site>
Cc:     Huacai Chen <chenhuacai@kernel.org>,
        WANG Xuerui <kernel@xen0n.name>, loongarch@lists.linux.dev,
        linux-kernel@vger.kernel.org, loongson-kernel@lists.loongnix.cn
Content-Type: text/plain; charset="UTF-8"
Content-Transfer-Encoding: 8BIT
Precedence: bulk

On Mon, Apr 10, 2023 at 8:54 PM Xi Ruoyao <xry111@xry111.site> wrote:
> Regarding these functions: we have -ffreestanding which is preventing
> the compiler from optimizing for e.g. "memcpy(a, b, 8);" into a simple
> ld.d/st.d pair.  A explicit compiler built-in usage like
>
>   #define memcpy(a, b, c) __builtin_memcpy(a, b, c)
>
> would allow the compiler to do such kind of optimization.   Will this
> improve the performance?

That's a good question. IIUC, the current compiler generates
inefficient code for constant-length memcpy, which may sacrifice
performance for compatibility, as not all hardware supports unaligned
memory access. We need a runtime CPU feature dispatch mechanism
similar to alternatives to improve the compiler. This is indeed a
problem that needs to be addressed.

Regards,
Rui