Re: [PATCH 2/2] Optimize compound_head() by avoiding a shared page flag

[Andrew Morton <akpm@linux-foundation.org>]

On Sat, 7 Apr 2007 18:32:04 -0700 (PDT) Christoph Lameter <clameter@sgi.com> wrote:

> On Sat, 7 Apr 2007, Andrew Morton wrote:
> 
> > > I just tried the approach that we discussed earlier and it was not 
> > > nice either.
> > 
> > We've discussed at least three approaches, so we don't know to what you refer.
> 
> Thats the approach of checking two flags at the same time. In that case 
> the compiler will generate and "and-immediate" and then a 
> "compare-immediate" one branch but .... Yuck.

Right.

        movl    (%ebx), %eax    # <variable>.flags, tmp399
        andl    $48, %eax       #, tmp399
        cmpl    $48, %eax       #, tmp399
        je      .L265   #,

what's "yuck" about that?

With the single page flag:

	movl	(%ebx), %eax	#* page.521, D.21940
	testb	$32, %al	#, D.21940
	jne	.L265	#,

So you're talking about saving one sole single silly solitary instruction.


> > Because I don't expect there will be much efficiency difference between the
> > above and the use of another page flag.
> 
> Then we end up with all these small efficiency differences in all 
> the code paths. I'd rather go for optimal performance in a frequently used 
> construct like this.

You can save that worrisome single instruction in the common case by putting the
handling of the uncommon compound pages out of line, as I indicated.