linux-kernel - Re: [PATCH] lib/int_sqrt.c: Optimize square root function

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]

Message-ID: <CA+55aFyC_koMROzCQTqzsKMSamHCmhna2bUc8UHW_WcRNO0bMg@mail.gmail.com>
Date:	Thu, 5 Feb 2015 10:20:29 -0800
From:	Linus Torvalds <torvalds@...ux-foundation.org>
To:	Anshul Garg <aksgarg1989@...il.com>
Cc:	Davidlohr Bueso <dave@...olabs.net>,
	Linux Kernel Mailing List <linux-kernel@...r.kernel.org>,
	"anshul.g@...sung.com" <anshul.g@...sung.com>
Subject: Re: [PATCH] lib/int_sqrt.c: Optimize square root function

On Tue, Feb 3, 2015 at 7:42 AM, Anshul Garg <aksgarg1989@...il.com> wrote:
>
> I have done profiling of int_sqrt function using perf tool for 10 times.
> For this purpose i have created a userspace program which uses sqrt function
> from 1 to a million.

Hmm. I did that too, and it doesn't improve things for me. In fact, it
makes it slower.

    [torvalds@i7 ~]$ gcc -Wall -O2 -DREDUCE int_sqrt.c ; time ./a.out

    real 0m2.098s
    user 0m2.095s
    sys 0m0.000s

    [torvalds@i7 ~]$ gcc -Wall -O2 int_sqrt.c ; time ./a.out

    real 0m1.886s
    user 0m1.883s
    sys 0m0.000s

and the profile shows that 35% of the time is spent on that branch
back of the initial reduction loop.

In contrast, my suggested "reduce just once" does seem to improve things:

    [torvalds@i7 ~]$ gcc -Wall -O2 -DONCE int_sqrt.c ; time ./a.out

    real 0m1.436s
    user 0m1.434s
    sys 0m0.000s

but it's kind of hacky.

NOTE! This probably depends a lot on microarchitecture details,
including very much branch predictor etc. And I didn't actually check
that it gives the right result, but I do think that this optimization
needs to be looked at more if we want to do it.

I was running this on an i7-4770S, fwiw.

Attached is the stupid test-program I used to do the above. Maybe I
did something wrong.

                         Linus

View attachment "int_sqrt.c" of type "text/x-csrc" (951 bytes)