Re: [PATCH] x86: Turn off GCC branch probability heuristics

From: Ingo Molnar <mingo@kernel.org>
To: Thomas Gleixner <tglx@linutronix.de>
Cc: Linus Torvalds <torvalds@linux-foundation.org>,
	Denys Vlasenko <dvlasenk@redhat.com>,
	Borislav Petkov <bp@alien8.de>,
	Tim Chen <tim.c.chen@linux.intel.com>,
	Andy Lutomirski <luto@amacapital.net>,
	Jason Low <jason.low2@hp.com>, Brian Gerst <brgerst@gmail.com>,
	Aswin Chandramouleeswaran <aswin@hp.com>,
	"Paul E. McKenney" <paulmck@linux.vnet.ibm.com>,
	Davidlohr Bueso <dave@stgolabs.net>,
	Peter Zijlstra <a.p.zijlstra@chello.nl>,
	"H. Peter Anvin" <hpa@zytor.com>,
	LKML <linux-kernel@vger.kernel.org>,
	Peter Zijlstra <peterz@infradead.org>
Subject: Re: [PATCH] x86: Turn off GCC branch probability heuristics
Date: Sun, 12 Apr 2015 10:07:04 +0200	[thread overview]
Message-ID: <20150412080704.GC9062@gmail.com> (raw)
In-Reply-To: <alpine.DEB.2.11.1504112038590.3845@nanos>

* Thomas Gleixner <tglx@linutronix.de> wrote:

> > How much of that 5% comes from code alignment? Or was this on 
> > *top* of the 1-byte alignment testt?
> 
> I thinks its just the no-guess one:
> 
>    text	       data       dec  patch           reduction
> 
> 7563475	    1781048  10302987   
> 
> 7192973	    1780024   9931461  no-guess	           -4.8%
> 
> 7354819	    1781048    958464  align-1		   -2.7%
> 
> 7192973	    1780024   9931461  no-guess + align-1  -4.8%
> 
> So with the no-guess applied the align-1 does not matter anymore.

What kernel config and GCC version did you use?

I used x86-64 defconfig+kvmconfig with GCC 4.9.1, and it gave me:

     text   filename                                        reduction
 12566383   vmlinux.expect=10 [==vanilla]
 11923529   vmlinux.-fno-guess-branch-probability           -5.4%
 11903663   vmlinux.align=1                                 -5.6%
 11646102   vmlinux.align=1+fno-guess-branch-probability    -7.9%

So there's still appears to be an additional -2.5% to be gained from 
turning off GCC branch heuristics.

x86 defconfig was derived from a distro kernel config and is still 
pretty close to what distros typically enable, so it's a good 
'typical' config to use for testing.

To double check that assumption I also tested a distro kernel .config 
and looked at vmlinux sizes:

     text   filename                                        reduction
  12612201  vmlinux.expect=10 [==vanilla]
  12107614  vmlinux.-fno-guess-branch-probability           -4.1%
  12021817  vmlinux.align=1                                 -4.9%
  11846771  vmlinux.align=1+fno-guess-branch-probability    -6.5%

this was cloned from a different major Linux distro than the one the 
original x86 defconfig was derived from - still the vmlinux sizes are 
pretty similar.

So x86 'make defconfig' measurements are pretty representative if you 
are looking for a quick, independent way to measure 'typical Linux 
distro' kernel characteristics.

( The only bigger difference is that FUNCTION_TRACER was turned on in 
  the distro config - which bloats function prologues a bit and thus 
  reduces the relative savings a bit. )

Thanks,

	Ingo