Re: [PATCH 2/3] netfilter : 3 patches to boost ip_tables performance

From: Willy Tarreau <willy@w.ods.org>
To: Eric Dumazet <dada1@cosmosbay.com>
Cc: Harald Welte <laforge@netfilter.org>,
	netdev@vger.kernel.org, netfilter-devel@lists.netfilter.org,
	linux-kernel@vger.kernel.org, Andi Kleen <ak@suse.de>
Subject: Re: [PATCH 2/3] netfilter : 3 patches to boost ip_tables performance
Date: Fri, 23 Sep 2005 06:02:34 +0200	[thread overview]
Message-ID: <20050923040234.GC595@alpha.home.local> (raw)
In-Reply-To: <4332AC2E.8000607@cosmosbay.com>

On Thu, Sep 22, 2005 at 03:05:50PM +0200, Eric Dumazet wrote:
(...) 
> It was necessary to get the best code with gcc-3.4.4 on i386 and 
> gcc-4.0.1 on x86_64
> 
> For example :
> 
> bool1 = FWINV(ret != 0, IPT_INV_VIA_OUT);
> if (bool1) {
> 
> gives a better code than :
> 
> if (FWINV(ret != 0, IPT_INV_VIA_OUT)) {
> 
> (one less conditional branch)
> 
> Dont ask me why, it is shocking but true :(

I also noticed many times that gcc's optimization of "if (complex condition)"
is rather poor and it's often better to put it in a variable before. I even
remember that if you use an intermediate variable, it can often generate a
CMOV instruction on processors which support it, while it produces cond tests
and jumps without the variable. Generally speaking, if you want fast code,
you have to write it as a long sequence of small instructions, just as if
you were writing assembly. As you said, shocking but true.

BTW, cheers for your optimizations !

Regards,
Willy