[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <CAHk-=wg2-j8zocUjurAeg_bimNz7C5h5HDEXKK6PxDmR+DaHRg@mail.gmail.com>
Date: Tue, 26 Jul 2022 11:36:21 -0700
From: Linus Torvalds <torvalds@...ux-foundation.org>
To: Yury Norov <yury.norov@...il.com>
Cc: Dennis Zhou <dennis@...nel.org>,
Guenter Roeck <linux@...ck-us.net>,
Russell King - ARM Linux <linux@...linux.org.uk>,
Catalin Marinas <catalin.marinas@....com>,
Linux Kernel Mailing List <linux-kernel@...r.kernel.org>,
Geert Uytterhoeven <geert@...ux-m68k.org>,
linux-m68k@...ts.linux-m68k.org
Subject: Re: Linux 5.19-rc8
On Tue, Jul 26, 2022 at 11:18 AM Yury Norov <yury.norov@...il.com> wrote:
>
> We have find_bit_benchmark to check how it works in practice. Would
> be great if someone with access to the hardware can share numbers.
Honestly, I doubt benchmarking find_bit in a loop is all that sensible.
These are helper functions that are probably seldom super-hot in
cache, and as with so many of these things, I suspect the cool-I$
numbers are the ones that matter most in real life.
When some filesystem ends up searching for a free block or similar, it
will probably have done other things before that that means that the
L1 I$ has been long flushed, and branch history is quite possibly
entirely gone too.
The same is quite possibly true for the bitmap itself in D$ too.
That said, looking at the x86 code generation (not only because I have
the build environment, but where I can actually read make sense of the
asm), the only thing that looks bad is the conditional bswap.
So the code gen looks fairly good,
Linus
Powered by blists - more mailing lists