Re: [patch] IBM HDAPS accelerometer driver, with probing.

[Alexey Dobriyan <adobriyan@gmail.com>]

On Sat, Aug 27, 2005 at 01:12:50AM -0500, Dmitry Torokhov wrote:
> On Saturday 27 August 2005 00:34, Arnaldo Carvalho de Melo wrote:
> > Em Fri, Aug 26, 2005 at 09:06:22PM -0700, Mitchell Blank Jr escreveu:
> > > Andi Kleen wrote:
> > > > - it doesn't seem to help that much on modern CPUs with good
> > > > branch prediction and big icaches anyways.
> > > 
> > > Really?  I would think that as pipelines get deeper (although that trend
> > > seems to have stopped, thankfully) and Icache-miss penalties get relatively
> > > larger we'd see unlikely() becoming MORE of a benefit, not less.  Storing
> > > the used part of a "hot" function in 1 Icacheline instead of 4 seems like
> > > an obvious win.
> > > 
> > > Personally I've never found unlikely() to be ugly; if anything I think
> > > it serves as a nice little human-readable comment about whats going on
> > > in the control-flow.  I guess I'm in the minority on that one, though.
> > 
> > Hey, even if unlikely was:
> > 
> > #define unlikely(x) (x)
> > 
> > I'd find it useful :-)
> >
> 
> Aside from annotating performance-critical sections what other purpose
> would it carry? It's not like you should not pay attention to teh code
> in these branches even if the are unlikely to be taken. So if code is
> not in hot path likely/unlikely just litter the code.
> 
> Btw, does it actually generate smaller code for constructs like
> 
> 	if (unlikely(blah))
> 		goto out;

Well, with my usual .config (-O2) and gcc-3.3.5-something it does:

text    data     bss     dec     hex filename
3614     303    1696    5613    15ed drivers/hwmon/hdaps.o
3678     303    1696    5677    162d drivers/hwmon/hdaps.o (unlikely()s removed)

Fortunately, there is -Os:

text    data     bss     dec     hex filename
3163     303    1696    5162    142a drivers/hwmon/hdaps.o
3163     303    1696    5162    142a drivers/hwmon/hdaps.o (unlikely()s removed)

See? The difference is 64 vs 451 bytes.

-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/