答复: 答复: [PATCH][nf-next] netfilter: replace modulo operation with bitwise AND

["Li,Rongqing" <lirongqing@xxxxxxxxx>]

> -----邮件原件-----
> 发件人: Florian Westphal [mailto:fw@xxxxxxxxx]
> 发送时间: 2019年2月25日 19:54
> 收件人: Li,Rongqing <lirongqing@xxxxxxxxx>
> 抄送: Florian Westphal <fw@xxxxxxxxx>; netfilter-devel@xxxxxxxxxxxxxxx
> 主题: Re: 答复: [PATCH][nf-next] netfilter: replace modulo operation with
> bitwise AND
> 
> Li,Rongqing <lirongqing@xxxxxxxxx> wrote:
> >
> >
> > > -----邮件原件-----
> > > 发件人: Florian Westphal [mailto:fw@xxxxxxxxx]
> > > 发送时间: 2019年2月25日 19:27
> > > 收件人: Li,Rongqing <lirongqing@xxxxxxxxx>
> > > 抄送: netfilter-devel@xxxxxxxxxxxxxxx
> > > 主题: Re: [PATCH][nf-next] netfilter: replace modulo operation with
> > > bitwise AND
> > >
> > > Li RongQing <lirongqing@xxxxxxxxx> wrote:
> > > > CONNTRACK_LOCKS is 1024 and power of 2, so modulo operations can
> > > > be replaced with AND (CONNTRACK_LOCKS - 1)
> > > >
> > > > and bitwise AND operation is quicker than module operation
> > >
> > > Uh.  What kind of compiler doesn't figure that out?!
> > >
> > > I would prefer to keep it as-is and let compiler do the optimization.
> >
> >
> > gcc version 7.3.0 (GCC)
> >
> >
> > main()
> > {
> >     int i=1000000000;
> >
> >          i= i % 1024;
> >
> >         return i;
> > }
> >
> > 00000000004004a7 <main>:
> >   4004a7:       55                      push   %rbp
> >   4004a8:       48 89 e5                mov    %rsp,%rbp
> >   4004ab:       c7 45 fc 00 ca 9a 3b    movl   $0x3b9aca00,-0x4(%rbp)
> >   4004b2:       8b 45 fc                mov    -0x4(%rbp),%eax
> >   4004b5:       99                      cltd
> >   4004b6:       c1 ea 16                shr    $0x16,%edx
> >   4004b9:       01 d0                   add    %edx,%eax
> >   4004bb:       25 ff 03 00 00          and    $0x3ff,%eax
> 
> & 1023.
> 
> With -O2 gcc emits same code regardless of % 1024 or & 1023.
> 
> > Similar patch:
> >
> > commit 1a1d74d378b13ad3f93e8975a0ade0980a49d28b
> > Author: Jakub Kicinski <jakub.kicinski@xxxxxxxxxxxxx>
> > Date:   Mon Oct 31 20:43:17 2016 +0000
> >
> >     nfp: use AND instead of modulo to get ring indexes
> 
> But in that case the value isn't known at compile time.
> 
> gcc can't know that '= reg1 % reg2' can be rewritten as 'reg1 & (reg2 - 1)'
> 
> But it can do it if reg2 is a known constant (and a power of 2).


Thanks, You are right

I think this patches maybe doing two thing:

1. make codes not depend on compiler optimization
2. remind to not change CONNTRACK_LOCKS to a value which is not power of 2

-RongQing