Re: [PATCH] arm: lib: implement aeabi_uldivmod via div64_u64_rem

From: Arnd Bergmann <arnd@kernel.org>
To: Nick Desaulniers <ndesaulniers@google.com>
Cc: clang-built-linux <llvm@lists.linux.dev>,
	Nathan Chancellor <nathan@kernel.org>,
	 Miguel Ojeda <ojeda@kernel.org>,
	Ard Biesheuval <ardb@kernel.org>, Gary Guo <gary@garyguo.net>,
	 Russell King <linux@armlinux.org.uk>,
	Linux ARM <linux-arm-kernel@lists.infradead.org>,
	 Linux Kernel Mailing List <linux-kernel@vger.kernel.org>
Subject: Re: [PATCH] arm: lib: implement aeabi_uldivmod via div64_u64_rem
Date: Sat, 16 Jul 2022 11:46:59 +0200	[thread overview]
Message-ID: <CAK8P3a0ky5w82=d2N_iZ2HYujeCDKJJqjT8P25-_+VpsZYWA_A@mail.gmail.com> (raw)
In-Reply-To: <20220716001616.4052225-1-ndesaulniers@google.com>

On Sat, Jul 16, 2022 at 2:16 AM Nick Desaulniers
<ndesaulniers@google.com> wrote:
>
> Compilers frequently need to defer 64b division to a libcall with this
> symbol name. It essentially is div64_u64_rem, just with a different
> signature. Kernel developers know to call div64_u64_rem, but compilers
> don't.
>
> Link: https://lore.kernel.org/lkml/20220524004156.0000790e@garyguo.net/
> Suggested-by: Gary Guo <gary@garyguo.net>
> Signed-off-by: Nick Desaulniers <ndesaulniers@google.com>

This has historically been strongly NAK'd, and I don't think that position
has changed in the meantime. A variable-argument 64-bit division is
really expensive, especially on 32-bit machines that lack a native
32-bit division instruction, and we don't want developers to accidentally
insert one in their driver code.

Explicitly calling one of the division helpers in linux/math64.h is the
established way for driver writers to declare that a particular division
cannot be turned into a cheaper operation and is never run in a
performance critical code path. The compiler of course cannot know
about either of those.

        Arnd